Avoid parsing upstream respose using Pydantic where possible

[adubovik]

See Mistral adapter:

def _response_to_dict(obj: ChatCompletionChunk | ChatCompletion) -> dict:
    return obj.to_dict(warnings=False)

async def chat_completion(
    *, request: dict, client: AsyncAzureOpenAI | AsyncOpenAI
) -> AsyncIterator[dict] | dict:
    response: (
        AsyncStream[ChatCompletionChunk] | ChatCompletion
    ) = await call_with_extra_body(client.chat.completions.create, request)

    if isinstance(response, AsyncStream):
        raw_stream = map_stream(_response_to_dict, response)
        return extract_reasoning_content(raw_stream)
    else:
        return extract_reasoning_content(_response_to_dict(response))

Here, openai library does parsing for the response, and then we convert the models back to JSON.
This round trip could be avoided completely if we use client.chat.completions.with_raw_response.create instead, which doesn't do any parsing.

Likewise for other adapters that do not require response parsing.