fix: handle FastAPI serialization for Message objects

[OiPunk]

Summary

• Fixes fastapi直接返回Message对象，序列化报错，BUG #347: BaseModelCompatibleDict.model_dump() override was being bypassed by FastAPI's serialization pipeline
• Added @model_serializer(mode='wrap') to BaseModelCompatibleDict that filters out None values at the Pydantic serialization level
• This ensures consistent exclude_none=True behavior whether serialized via direct model_dump(), Pydantic's TypeAdapter, or FastAPI's jsonable_encoder

Problem

The overridden model_dump method only sets exclude_none=True when the caller doesn't pass the parameter. However:

1. FastAPI's jsonable_encoder calls model_dump(exclude_none=False) explicitly, overriding the intended default
2. Pydantic's TypeAdapter (used by FastAPI for response_model validation) bypasses model_dump() entirely

This causes None-valued fields (name, function_call, extra, etc.) to appear as null in API responses.

Solution

Add a @model_serializer(mode='wrap') that operates at the Pydantic core serialization level, ensuring None values are filtered regardless of the call path. The existing model_dump/model_dump_json overrides are preserved for backward compatibility.

Test plan

• Message.model_dump() excludes None fields
• Message.model_dump_json() excludes None fields
• ContentItem.get_type_and_value() still works (depends on single-item dict)
• TypeAdapter(Message).dump_python() excludes None fields
• TypeAdapter(Message).dump_json() contains no null values
• FastAPI endpoints return clean JSON without null fields
• Dict-like access (__getitem__, __setitem__, get) still works
• Added unit tests in tests/llm/test_schema.py

[OiPunk]

Closing — shifting focus to AI Agent core projects. Thank you for your time!