fix: accept OpenAI-compatible chat completions request shape

[aviweit]

fix: accept OpenAI-compatible chat completions request shape

Problem

POST /chat/completions returned 422 Unprocessable Entity for any
client following the full OpenAI chat completions spec, including
OpenHands. Two distinct validation failures were occurring:

1. content as a list of content blocks — OpenAI spec allows
   content to be [{"type": "text", "text": "..."}] (multimodal
   format). ChatMessage.content: str rejected this with
   Input should be a valid string.

2. content: null on assistant tool-call messages — When an
   assistant message contains tool_calls, OpenAI sets content: null.
   Same rejection as above.

3. Unknown fields — OpenAI clients send stream, top_p, stop,
   frequency_penalty, presence_penalty, etc. Pydantic's default
   strict mode rejected these with 422.

Changes (fast_api/api_server.py)

What	Before	After
ChatMessage.content type	str	Optional[Union[str, List[ContentBlock]]]
content: null	422	accepted, treated as ""
content: [{"type":"text","text":"..."}]	422	accepted, text parts joined
Unknown fields on ChatMessage	422	silently ignored
Unknown fields on ChatRequest	422	silently ignored

New ContentBlock model parses {type, text} blocks. New
_content_as_str() method flattens any content shape to a plain string
before passing to LangChain.

Backward compatibility

Fully backward compatible. Existing callers sending content as a plain
string hit the same str branch of the Union as before — no behaviour
change. The fix only widens what is accepted.

Testing

Verified with curl against the running server:

• plain string content → 200 ✅
• content: null → 200 ✅
• content: [{"type":"text","text":"..."}] → 200 ✅
• extra fields (stream, top_p, …) → 200 ✅
• end-to-end OpenHands bench run via BenchFlow → agent reaches LLM, no more 422s ✅