[Bugfix][Router] Preserve full backend model metadata in /v1/models

[yzhan1]

Summary

This fixes #892 by preserving extra fields returned by backend /v1/models responses instead of rebuilding a minimal router-only model card.

The change stores unknown backend model fields in ModelInfo.extra_fields and serializes them back out from the router's /v1/models endpoint. That keeps metadata such as max_model_len and permission available to clients.

Validation

• uv run --group test python -m pytest -q src/tests/test_main_router_models.py
• uv run --group test python -m pytest -q src/tests
• Live check against a real Dockerized vLLM backend (vllm/vllm-openai-cpu:latest-arm64):
  • backend /v1/models included max_model_len and permission
  • fork/main router output dropped those fields
  • this branch preserved them (base_missing=['max_model_len', 'permission'], fix_missing=[])

[gemini-code-assist]

Code Review

This pull request enables the preservation of extra metadata fields in ModelInfo and ensures they are returned by the models endpoint. While the implementation includes new tests, feedback highlights that instantiating ModelCard with these extra fields may cause excessive warning logs due to existing model validators. Additionally, it is recommended to move the known_fields set to a class-level constant to improve performance.