feat: Add SpanHandler customization to langfuse integration

[vblagoje]

Why:

Enhances the Langfuse integration by adding a flexible mechanism for custom span processing, allowing users to define how spans should be handled within their specific LLM components.

• fixes Add SpanHandler abstraction to enable custom LLM tracing in Langfuse #1310

What:

• Introduced SpanHandler, a callable type for custom span processing
• Added span handler support to LangfuseConnector and LangfuseTracer
• Default span processing behavior is preserved through default_span_handler

How can it be used:

from haystack import Pipeline
from haystack_integrations.components.connectors.langfuse import LangfuseConnector
from haystack_integrations.components.generators.custom import CustomChatGenerator

def custom_span_handler(span, component_type):
    if component_type == "CustomChatGenerator":
        output = span._data.get("haystack.component.output", {})
        span._span.update(
            usage=output.get("usage"),
            model="custom-llm-model",
            custom_metadata=output.get("extra_info")
        )

# Add to your pipeline
pipe = Pipeline()
pipe.add_component(
    "tracer",
    LangfuseConnector(
        name="Custom LLM Pipeline",
        span_handler=custom_span_handler
    )
)
pipe.add_component("llm", CustomChatGenerator())

How did you test it:

• Unit tests for span handler serialization/deserialization
• Integration tests with custom span handler
• Verified traces appear correctly in Langfuse UI

Notes for the reviewer:

The callable approach was chosen over a class-based one for simplicity and better serialization support. The handler has access to both span data and component type, making it flexible enough for custom LLM integrations.