`ParsedChatCompletion` type information lost during response conversion

[bilelomrani1]

Description

When response_format is set to a Pydantic model, the OpenAI provider correctly calls client.chat.completions.parse(), which returns an openai.ParsedChatCompletion[T] with a typed message.parsed: T field containing a validated Pydantic model instance.

However, _convert_chat_completion round-trips the response through model_dump() → ChatCompletion.model_validate(), which serializes the Pydantic instance on message.parsed into a plain dict. The data is preserved (thanks to extra="allow" on ChatCompletionMessage), but the typed model is gone.

Example

from pydantic import BaseModel

class CalendarEvent(BaseModel):
    name: str
    date: str
    participants: list[str]

client = AnyLLM.create("openai")
response = await client.acompletion(
    model="gpt-4o",
    messages=[{"role": "user", "content": "List a birthday party event"}],
    response_format=CalendarEvent,
)

parsed = response.choices[0].message.parsed

# Expected: CalendarEvent(name='...', date='...', participants=['...'])
# Actual:   {'name': '...', 'date': '...', 'participants': ['...']}

Would it be possible to either preserve the ParsedChatCompletion type through the conversion, or expose a dedicated return path for structured output that keeps the typed parsed field?

[njbrake]

I liked your PR, let's get rid of the OpenAI parse/create distinction, which makes it more straightforward