feat: make scheduler openai agent compatible

[gufengc]

Summary

• Add /v1/models to the scheduler OpenAI gateway.
• Keep /v1/chat/completions client fields pass-through while stripping/overriding Parallax-internal routing fields.
• Preserve downstream non-streaming HTTP status/content type over the RPC boundary and return OpenAI-shaped scheduler errors.
• Cover model listing, agent field pass-through, OpenAI errors, and downstream status preservation in tests.

Validation

• pre-commit run --all-files
• .venv/bin/python -m pytest (133 passed, 5 skipped)
• Manual README flow: parallax run -m Qwen/Qwen3-0.6B -n 1, parallax join --use-hfcache, scheduler /v1/models, streaming chat completion, tool call request, and follow-up role: "tool" response all succeeded.