docs(walkthroughs): add scenario walkthroughs and align CR manifests

[thxCode]

What type of PR is this?

/kind documentation
/area worker

What this PR does / why we need it:

• Adds 4 new scenario-driven walkthroughs derived from operator feature deployment scenarios:
  • docs/walkthroughs/rdma/network-endpoints.md: requesting and verifying RDMA endpoints and NUMA alignment.
  • docs/walkthroughs/topology/hierarchical-placement.md: placing serving replicas across availability zones and physical racks.
  • docs/walkthroughs/model-deployment/prefill-decode.md: deploying prefill and decode roles fronted by router.
  • docs/walkthroughs/model-deployment/multi-host.md: scaling serving replicas across multiple GPU nodes.
• Refactors existing walkthroughs to prefer manifest diffs over redundant YAML declarations.
• Standardizes Kubernetes custom resource manifests across all walkthroughs and module pages with complete apiVersion and kind declarations.
• Updates Kueue scheduling resources to v1beta2.
• Aligns sidebar menu labels, page H1 headings, and cross-reference links across module guide landing pages.
• Updates troubleshooting runbooks with verified condition reasons, container logs, and UID ownership.

Validation:

• bash .agents/skills/gpustack-operator-docs/scripts/check-docs.sh --report passed (66 docs pages checked, 0 errors, heading/TOC/paragraph length constraints satisfied).
• make lint docs passed (specs verified, Hugo site built 151 pages, 0 broken internal links, agent-index.py verified llms.txt and exported Markdown consistency).

Which issue(s) this PR links to:

NONE

Special notes for your reviewer:

• All new walkthrough pages are indexed in docs/README.md, docs/walkthroughs/_index.md, and site/data/navigation.yaml.
• Verified that llms.txt accurately includes all new walkthroughs and mirrors docs/README.md#All pages.

Does this PR introduce a user-facing change?

Add scenario-driven walkthroughs for RDMA network endpoints, hierarchical placement, prefill/decode serving, and multi-host serving, and align CR YAML manifests.

[gpustack-code-review]

✅ OpenCodeReview: Review complete: 0 finding(s) across 3 selected item(s).

Reviewed e2d6bb6..2a3bfa3 only; earlier commits in this PR were reviewed in a previous run.

[thxCode]

Review Feedback Triage & Resolution Summary

All 13 findings reported by OpenCodeReview have been triaged and addressed:

Inline Review Threads

1. docs/modules/rdma/operations.md:274 (Base image digest): Kept as-is by design. The placeholder digest was intentionally omitted in this Dockerfile template to keep the snippet clean and directly usable without manual placeholder edits. (Thread left open).
2. docs/walkthroughs/kv-cache/shared-cache.md:363 (Condition reasons in Troubleshooting table): Verified against controller source. All condition reasons (BackendResolved, MemberCrashLooping, MemberInitCrashLooping, SegmentsShort, PreexistingContent, ZeroGranted, DomainClaimedByMultipleBindings) are exact string constants emitted by the operator reconcilers in pkg/worker/controllers/worker/. (Thread resolved).
3. docs/walkthroughs/model-deployment/multi-host.md:82 (vLLM multi-node command clarification): Fixed. Clarified that the snippet illustrates operator coordinate injection (GPUSTACK_* envs) for custom entrypoints, and noted that production multi-node vLLM deployments wrap Ray (--distributed-executor-backend ray) or torchrun (external_launcher). (Thread resolved).
4. docs/walkthroughs/rdma/network-endpoints.md:29 (RDMA resource keys table): Fixed. Added cross-links to the owning Network Topology module and its resource keys section. (Thread resolved).
5. docs/walkthroughs/rdma/network-endpoints.md:77 (Injected devices list duplication): Fixed. Replaced the duplicated device list with a cross-link to Injected devices in the RDMA operations module. (Thread resolved).
6. docs/walkthroughs/rdma/network-endpoints.md:99 (TopologyManager kubelet config): Fixed. Added cross-link to Enabling NUMA alignment on the kubelet in the RDMA operations module. (Thread resolved).
7. docs/walkthroughs/rdma/network-endpoints.md:180 (Fabric privileges description): Fixed. Replaced duplicated privileges description with a cross-link to The members in the KV cache backend module. (Thread resolved).
8. site/data/navigation.yaml:34 (Navigation ordering for sub-sections): Fixed. Restored per-subsection order keys and added subsection labels so Hugo resolves and orders sub-sections properly. (Thread resolved).

Summary-Routed Findings

9. docs/walkthroughs/devices/scheduling.md:335 (NodeFeature naming reference): Fixed. Added cross-link to Device Scheduling module.
10. docs/walkthroughs/rdma/network-endpoints.md:145 (1-vs-N interface mapping): Fixed. Added cross-link to Model Deployment configuration.
11. docs/walkthroughs/topology/hierarchical-placement.md:195 (Topology operations reference): Kept as-is. Preserved walkthrough self-containment for runnable operator workflows while maintaining link to operations.md#change-a-live-hierarchy.
12. docs/walkthroughs/model-delivery/prefetch.md:211 (Prefetch condition semantics): Kept as-is. Retained troubleshooting table self-containment for runnable operator workflows.
13. docs/walkthroughs/model-deployment/external-dp.md:328 (Validation record runner/router image digests): Fixed. Added observed runner digest (gpustack/runner:cuda13.0-vllm0.29.0@sha256:1c8267252069b165fb3e19875f5697669ba8e7525fc6eb142b781165ad5730a3) and router digest (gpustack/llm-router:dev@sha256:98e70d944c6c2e3a1523ecddcb2d2c161eb32d786db9fb42fa103dbf072bbd92) to the Validation Record.

All documentation checks (check-docs.sh --report and make lint docs) pass with zero errors.

[thxCode]

Review Feedback Triage & Resolution Summary (Round 2)

All 4 findings reported in the latest review pass have been resolved:

1. docs/walkthroughs/model-deployment/external-dp.md:328 (Validation record image digests): Fixed in commit af3413d4. Removed the sha256 digests for runner and router in favor of clean version tags matching the deployment manifest (gpustack/runner:cuda13.0-vllm0.29.0 and gpustack/llm-router:v0.2.0). (Thread resolved).
2. docs/walkthroughs/model-deployment/multi-host.md:77 (vLLM distributed backend flag): Fixed. Updated the example command from --distributed-executor-backend mp to --distributed-executor-backend ray to ensure valid multi-node configuration. (Thread resolved).
3. docs/walkthroughs/model-deployment/multi-host.md:202 (Single-role QuotaReserved condition reason): Fixed. Replaced QuotaReserved=False/Parked with QuotaReserved=False/Pending in the troubleshooting table, accurately reflecting single-role quota wait behavior. (Thread resolved).
4. docs/walkthroughs/topology/hierarchical-placement.md:214 (Duplicated failure surfaces in troubleshooting): Fixed. Trimmed redundant general queue and migration rows from the troubleshooting table and added a direct cross-link to Failure Surfaces in the Topology scheduling module. (Thread resolved).

All verification checks (check-docs.sh --report and make lint docs) pass with zero errors.

[thxCode]

PR Review Round 3 Summary

All findings from the latest review round have been addressed:

• docs/walkthroughs/rdma/network-endpoints.md:
  • Replaced the duplicated RDMA resource keys table with a concise prose reference pointing to the RDMA network topology module.
  • Aligned the troubleshooting entry for loss of feature.gpustack.ai/rdma.capable to state that the label is removed when every endpoint is unusable (fails or lacks a bound RDMA device), with a cross-link to Network Topology.
• docs/walkthroughs/topology/hierarchical-placement.md:
  • Trimmed the duplicated HoldAndDrain sequence explanation to contrast revision-only changes with level changes and linked directly to the Topology operations guide.
  • Retained only walkthrough-specific symptoms in the troubleshooting table and cross-linked general failure surfaces to the Topology scheduling module.
• docs/walkthroughs/model-deployment/multi-host.md:
  • Aligned the multi-host container command to initialize Ray head on member m0 and join subordinate members via ray start --address=... --block before running vLLM with --distributed-executor-backend ray.

All documentation checks and linters passed cleanly (check-docs.sh, make lint docs).