feat: add multi-prefill-pool support for modality-based routing

[linzebing]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Enable independent scaling of text and perception (multimodal) prefill pods while sharing a single decode pool. Requests are auto-routed to the correct prefill pool based on modality detection (image_url content parts → perception, otherwise → text).

Key changes:

• Add detect_prefill_pool() / detect_prefill_pool_from_json() in spec.rs
• Add parse_prefill_selectors() for named pool CLI selectors (e.g. --prefill-selector=text:app=text-prefill)
• Change DiscoveryConfig.prefill_selector to prefill_selectors map
• Add WorkerRegistry.get_prefill_workers_by_pool() with label filtering
• Update K8s service discovery PodType::Prefill to carry pool name
• Add pool-aware routing in VllmPDRouter (route_chat, route_completion, route_transparent)
• Add prefill_pool_routed_total Prometheus metric with pool label
• Add 33 unit tests covering all new functionality
• Backward compatible: no pool prefix defaults to "default" pool

Test Plan

Tested by deploying multi-prefill on k8s.

Test Result

Both text and perception evals passed, and we verified routing worked correctly.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results