fix(api): forward extra_body.chat_template_kwargs on /v1/messages

[sufubao]

Summary

• The /v1/messages request translator strips extra_body before building the ChatCompletionRequest, so clients cannot reach LightLLM fields that Anthropic's schema does not expose — most notably chat_template_kwargs. On engines where thinking defaults to on (Qwen3, DeepSeek), this means Anthropic-protocol callers have no way to disable thinking; with short max_tokens the response degenerates to content: [] because every output token went into the <think> block.
• Merge any dict-valued extra_body into the translated openai_dict before the _UNKNOWN_FIELDS cleanup, with setdefault so fields produced by the Anthropic→OpenAI translation keep precedence. Unknown keys fall off naturally in ChatCompletionRequest(**chat_dict) since the Pydantic default is extra='ignore'.
• Usage example for clients:

  {
    \"model\": \"...\",
    \"max_tokens\": 512,
    \"messages\": [...],
    \"extra_body\": { \"chat_template_kwargs\": { \"enable_thinking\": false } }
  }

Test plan

• pytest test/test_api/test_anthropic_extra_body.py — 5 new unit tests covering: single field, multi-field, top-level field beats extra_body duplicate, missing extra_body is no-op, non-dict extra_body is ignored.
• black + flake8 pre-commit hooks pass.
• Manual: send an Anthropic request with extra_body.chat_template_kwargs.enable_thinking=false to a Qwen3 deployment and verify content is non-empty.

[gemini-code-assist]

Code Review

This pull request introduces support for the extra_body field in the Anthropic-to-OpenAI request translation, enabling the forwarding of LightLLM-specific parameters such as chat_template_kwargs. The logic correctly prioritizes existing translated fields and ignores non-dictionary extra_body inputs. Comprehensive unit tests have been included to validate field forwarding, precedence rules, and edge cases. I have no feedback to provide.