feat(router): pluggable load-balancing strategies + live dashboard sw…

[milk333445]

…itch

Make request routing a swappable policy instead of a hard-coded least-load selector. Adds a strategy registry with 7 strategies — least_load (default, unchanged behaviour), round_robin, random, least_inflight, p2c, and session_affinity / prefix_affinity (sticky routing with a load escape valve for multi-turn chat & shared-prompt cache reuse).

Router:

• routing_strategies.py: SelectContext + registry + score_instance (extracted from the old selector, so least_load is byte-for-byte identical); failover, in-flight accounting, and per-backend cooldown stay in the proxy and apply to every strategy.
• backend_selector.select_instance_least_load kept as a thin wrapper (compat).
• Strategy chosen per-group (model_config.routing_strategy) > global env (LLMOPS_ROUTING_STRATEGY) > default; session/prefix keys extracted in the proxy.
• GET/POST /routing to read + hot-swap the global strategy (no reload), exposed via nginx; dashboard Traffic page gets a selector + a per-strategy help card.

Frontend:

• routing strategy is a first-class field in the add/edit dialog and the detail drawer, kept out of the raw vLLM-param list (shared routingStrategies.ts).

Fixes:

• launchers: never pass router-only keys (routing_strategy) to vllm serve, which errored on the unknown argument.
• AddModelDialog: a null lora_modules from the config no longer leaks into the param list and gets re-submitted as "" (failed list validation).

Tests: router unit suite (strategies + /routing endpoint) and a launcher regression; frontend type-check clean.