Improve Langfuse Adapter

[xzrderek]

here's an example of the new pattern

def custom_converter(trace: TraceWithFullDetails, include_tool_calls: bool, span_name: Optional[str]) -> Optional[EvaluationRow]:
    messages = []

    if span_name:  # Look for a generation tied to a span name
        try:
            gen = find_final_generation_in_span(trace, span_name)
            if not gen:
                return None

            tools = []

            if gen.input:
                for d in gen.input:
                    if d["role"] == "tool":
                        tools.append(d["content"])
                    else:
                        messages.append(dict_to_message(d, include_tool_calls))
            if gen.output:
                messages.append(dict_to_message(gen.output, include_tool_calls))
            
            return EvaluationRow(
                messages=messages,
                tools=tools,
                input_metadata=InputMetadata(
                    session_data={
                        "langfuse_trace_id": trace.id,  # Store the trace ID here
                    }
                ),
            )
            
        except Exception as e:
            logger.error("Failed to extract messages from span '%s' in trace %s: %s", span_name, trace.id, e)
            return None
    else:
        logger.error("No span name provided")
        return None

@pytest.mark.asyncio
@evaluation_test(
    input_rows=[
        adapter.get_evaluation_rows(
            name="Subagent 1",
            limit=3,
            span_name="judge",
            environment="production",
            converter=custom_converter,
        )

[dphuang2]

nit: make this a protocol

[dphuang2]

small nit but LGTM