feat: fleet storage tiers and model placement view

[yr75myjnqx-tech]

Summary

Brings the multi-tier model storage concept to sparkDash for DGX Spark fleets: a read-only Storage view that classifies every Spark's disks into storage tiers, inventories the model weight files on the fleet, and shows where each model lives — both as a local copy and as served across the CX7/ConnectX fabric.

This is fully self-contained: tier classification and model scanning run inside the existing SystemCollector (local + SSH), with no external registry or service dependency.

What's in the PR

Tier classification

Every mounted disk is labeled a tier:

• Hot — root NVMe (active working copy)
• Warm — other real local disks
• Cold — NAS / network mounts (/mnt/modelshelf, /media, /Volumes, /mnt, or any cifs/smb/smb3/nfs/nfs4 filesystem)

Per-Spark tierPaths in config/sparks.json override the heuristic per mount.

Model inventory

The collector scans the configured model directories (grouped per tier via optional per-Spark modelDirs) for weight files — .safetensors, .gguf, .bin, .pt, .pth, .ckpt — and reports each model by name, size, and tier. New models metrics domain, polled via POLL_INTERVAL_MODELS (default 30 s). Each scanned root is inspected one level deep; modelDirs should point at the directory whose direct children are the model folders (for HF-style nesting, the outer family dir, e.g. ~/models/hf).

Dual placement (resident + fabric)

A model is resident on a Spark holding a local copy. A peer that has the model loaded in its LLM probe — matched by model name — serves it to the fleet over the CX7/ConnectX fabric, so that Spark shows the model as placed over the fabric even with no local copy. This mirrors the storage intent: a "hot" model is available to the fleet either from local disk or from a peer's memory across the fabric.

UI

• New reserved Storage tab (fleet level), plus per-Spark tier cards and a model placement table
• Each Spark's Storage panel shows a Hot/Warm/Cold badge per disk

Verification

• npm test — 94 passing (incl. new tierClassify and collectModels unit tests)
• npm run typecheck and npm run build — clean
• Live smoke on a DGX Spark (Gx10) — the real mount table classifies correctly (root NVMe → Hot, cifs NAS /mnt/macmini → Cold, other local/snap disks → Warm) and collectModels correctly detected the resident DeepSeek-V4-Flash-0731 (155.4 GiB) from a modelDirs root

Notes for review

• Read-only and additive: no eviction, promotion, or write actions; no changes to existing storage/liveness behavior
• Out of scope on purpose: warm-tier USB config UI, eviction/promotion actions, activity history, non-DGX support