feat(huggingface): expose reasoning_content in AIMessageChunk

[dip9811111]

This PR enhances the HuggingFace integration by including reasoning_content extracted from the model response in AIMessageChunk.additional_kwargs.

Previously, AIMessageChunk did not expose the "reasoning" or "reasoning_content" field returned by HuggingFace models, making it inaccessible to downstream LangChain components.
This change ensures that any "reasoning" or "reasoning_content" value in the provider response is propagated and available to users.

The updated logic inside the HuggingFace formatter now includes:

def _convert_chunk_to_message_chunk(
    chunk: Mapping[str, Any], default_class: type[BaseMessageChunk]
) -> BaseMessageChunk:

    if hasattr(chunk, "choices"):
        choice_obj = chunk.choices[0]
        delta_obj = choice_obj.delta
        reasoning_content_value = getattr(delta_obj, "reasoning_content", None)
    else:
        reasoning_content_value = None

    choice = chunk["choices"][0]
    _dict = choice["delta"]

    [...]

    if role == "assistant" or default_class == AIMessageChunk:

        reasoning = _dict.get("reasoning") or reasoning_content_value
        additional_kwargs["reasoning_content"] = reasoning

        if usage := chunk.get("usage"):
            input_tokens = usage.get("prompt_tokens", 0)
            output_tokens = usage.get("completion_tokens", 0)
            usage_metadata = {
                "input_tokens": input_tokens,
                "output_tokens": output_tokens,
                "total_tokens": usage.get("total_tokens", input_tokens + output_tokens),
            }
        else:
            usage_metadata = None
        return AIMessageChunk(
            content=content,
            additional_kwargs=additional_kwargs,
            tool_call_chunks=tool_call_chunks,
            usage_metadata=usage_metadata,  # type: ignore[arg-type]
        )
    [...]

This makes reasoning traces accessible to agents, logging utilities, tracing tools, and any consumer of AIMessageChunk.

Issue

(#34011)

Dependencies

No new dependencies required.