chore(closes OPEN-10731): async LangChain callback handler tool call …

[gustavocidornelas]

Pull Request

Summary

Fix the LangChain callback handler so ToolMessage.artifact (returned by @tool(response_format="content_and_artifact")) is captured on the trace instead of being silently dropped. When the artifact is a list of LangChain Document instances, the page-content strings are also promoted to the trace-level
context reserved column, matching the existing retriever behavior.

Changes

• Add _extract_tool_output helper that handles three shapes LangChain may pass to on_tool_end: ToolMessage
  with .content/.artifact, raw (content, artifact) tuple, and plain string (backward-compat).
• Add _is_document_list predicate (non-empty sequence whose elements all have page_content).
• Rewrite _handle_tool_end to store the artifact in step.metadata["artifact"] and, when it's a document list,
  populate trace-level context. Step output stays a string so downstream agent loops feeding step.output back to
  the LLM aren't disturbed.
• Add a run_id → trace map on OpenlayerHandlerMixin (maintained in _start_step/_end_step on both sync and
  async handlers) and a _find_trace helper so context promotion works in pure-callback mode (no outer tracer.trace()
  wrapper).
• Wire the new _find_trace lookup into _handle_retriever_end and the _handle_chain_end source_documents branch so those paths also populate context in pure-callback mode.
• New test file tests/test_langchain_callback.py with 6 cases (plain-string, tuple, ToolMessage with dict
  artifact, document-list promotion sync + async, end-to-end metadata serialization).
• Example script examples/tracing/langchain/tool_artifact_example.py reproducing the reporter's scenario.

Context

OPEN-10731: Async LangChain callback handler tool call improvements

Testing

• Manual testing

Monitoring

• No expected impact