AugmentedLLMs should utilize structured responses

[strawgate]

Similar to #221

There are a number of places where AugmentedLLMs are first producing text responses and then translating to structured responses.

e.x. in the Orchestrator

    async def generate_structured(
        self,
        message: str | MessageParamT | List[MessageParamT],
        response_model: Type[ModelT],
        request_params: RequestParams | None = None,
    ) -> ModelT:
        """Request a structured LLM generation and return the result as a Pydantic model."""
        params = self.get_request_params(request_params)
        result_str = await self.generate_str(message=message, request_params=params)

        llm = self.llm_factory(
            agent=Agent(
                name="Structured Output",
                instruction="Produce a structured output given a message",
            )
        )

        structured_result = await llm.generate_structured(
            message=result_str,
            response_model=response_model,
            request_params=params,
        )

        return structured_result

We should refactor AugmentedLLMs to use structured requests internally and eliminate anywhere we are gathering a string and pushing it through a completion to structure it

[strawgate]

If there is agreement on this I'd be happy to offer a PR for the GenAI LLM and the Orchestrator

[strawgate]

Actually looking at this again, I think this flow is probably broken?

The original generation produces a plan result, and the original text completion is unaware of the desired response model, but an llm later tries to map it to the provided response model?