feat(router): route Anthropic Messages through virtual models

[m199369309]

Summary

• route Anthropic Messages requests for Token Router virtual model UIDs through the existing OpenAI-compatible Router data plane
• reuse the canonical Anthropic request/response/SSE adapter from feat(api): add unified Anthropic Messages protocol adapter #5389
• convert Router OpenAI errors and streams back into native Anthropic protocol responses
• keep external Anthropic x-api-key/Bearer credentials at the Supervisor boundary; only a configured internal Router token is forwarded
• close upstream responses and per-request HTTP clients on success, error, disconnect, and stream completion

Dependencies

• Depends on feat(router): add token-aware virtual model routing #5375
• Depends on feat(api): add unified Anthropic Messages protocol adapter #5389

This branch intentionally contains both dependency histories so it can be tested before they merge. After the dependencies merge, it will be rebased onto the latest official main so this PR's diff contains only the Anthropic Token Router integration.

Tests

• pytest -q xinference/api/tests/test_token_router_dispatch.py -k anthropic
• pytest -q xinference/api/tests/test_anthropic_messages.py xinference/api/tests/test_anthropic_protocol.py
• pre-commit run --files xinference/api/restful_api.py xinference/api/tests/test_anthropic_messages.py xinference/api/tests/test_token_router_dispatch.py

[gemini-code-assist]

Code Review

This pull request introduces an independent, token-aware router process for DeepSeek-V4 on Xinference, adding new control-plane REST endpoints, schemas, and an Anthropic Messages protocol adapter to integrate token-aware routing with the existing RESTful API. Feedback on these changes highlights critical startup issues where asyncio.Condition and asyncio.Lock are instantiated outside of an active event loop, which will cause immediate crashes. Additionally, a performance concern was raised regarding the creation of a new httpx.AsyncClient instance for every request, which should be replaced with a shared client to avoid latency and socket exhaustion.

[m199369309]

Additional CI fix included in b6d2d4c:

• Normalize request header names in _request_credential() so both Starlette Headers and plain mapping-based test requests preserve HTTP case-insensitive header semantics.
• This fixes the two failing advanced-auth tests for unauthorized and invalid API keys.

Validation:

• 45 passed across the focused router, tokenization, Anthropic Messages, dispatch, and advanced-auth suites.
• pre-commit run --files xinference/api/restful_api.py xinference/api/tests/test_token_router_dispatch.py passed all hooks.