feat(router): add token-aware virtual model routing

[m199369309]

Summary

• add a native xinference-router data-plane process with process-isolated tokenization, admission control, typed first-match routing rules, and HTTP/SSE proxying
• add V1 short/long compatibility plus V2 dynamic 1-16 backend configuration for the llm_chat route profile
• add Supervisor-backed Router configuration/runtime registries, virtual model discovery, OpenAI chat dispatch, internal runtime registration, heartbeat, config polling, and acknowledgement APIs
• protect existing deployments with the opt-in XINFERENCE_TOKEN_ROUTER_ENABLED feature gate (disabled by default)
• add Router management scopes, audit coverage, API/runtime/configuration tests, and CLI packaging

Validation

• focused Router suite: 98 passed
• pre-commit run --files <modified files>

Follow-ups

Tokenizer asset management, bundled assets, Web UI, Agent orchestration, and monitoring are intentionally split into dependent PRs.

[gemini-code-assist]

Code Review

This pull request introduces an independent, token-aware router for DeepSeek-V4 on Xinference, adding a persistent SQLite-backed configuration store, an in-memory runtime registry, and new REST endpoints for router management and internal runtime coordination. It also integrates a proxy layer into the restful API to route chat completions dynamically based on token budgets or typed rules. The review feedback highlights critical improvement opportunities: avoiding resource leaks on early stream disconnects by using Starlette's BackgroundTask instead of generator finally blocks, reusing a shared httpx.AsyncClient to prevent connection overhead, replacing a slow copy.deepcopy with a shallow copy during message normalization, and fixing a string concatenation typo ("t" "e").

[qinxuye]

Found several blocking correctness and lifecycle issues in the current head.

[qinxuye]

LGTM