fix(openai,openai-agents): Serialize response output items with model_dump()

[fbartho]

Description

OpenAIIntegration (Responses API) and OpenAIAgentsIntegration serialize tool calls and non-text message content with the deprecated Pydantic v2 dict() method, so every response containing a function call or a refusal emits PydanticDeprecatedSince20: The `dict` method is deprecated; use `model_dump` instead. when outputs are recorded.

Changes:

• openai: replace the four dict() calls in _set_common_output_data with model_dump().
• openai-agents: replace the two dict() calls in _set_output_data with model_dump().
• Add test_responses_api_output_serialization_emits_no_pydantic_deprecation_warning (openai, both the data_collection and send_default_pii paths) and test_response_output_serialization_emits_no_pydantic_deprecation_warning (openai-agents), which assert no PydanticDeprecatedSince20 is emitted.

On Pydantic v1: openai's BaseModel defines model_dump() when Pydantic v1 is installed, delegating to dict(), in every version from 1.0.0, the integration's minimum. The Chat Completions path in the same function already calls choice.message.model_dump() on that basis. openai-agents requires pydantic>=2.10.

https://github.com/openai/openai-python/blob/becc1d20eed83c1b8d85e15dc131a372d9dc7813/src/openai/_models.py#L296-L372

The tox matrix pins Pydantic v2, so I also ran tests/integrations/openai/test_openai.py against openai 3.29.0 with pydantic 1.10.26: everything passes, and the new test skips.

Issues

Closes #7827

[fbartho]

Sorry, I didn't realize you had already fixed the main issue I was experiencing