fix(langchain): capture structured output tool calls

[yaodong-shen]

📥 Pull Request

📘 Description

Fixes #1263.

LangChain's with_structured_output() returns an AIMessage whose text is often empty while the structured result is stored in tool_calls. The callback previously serialized only generation text, so AgentOps traces could not render structured outputs correctly.

This change:

• emits completions through AgentOps' indexed gen_ai.completion.{i} semantic attributes;
• captures normalized LangChain AIMessage.tool_calls and the legacy OpenAI-shaped additional_kwargs.tool_calls form;
• preserves text completions, roles, finish reasons, tool metadata, and JSON arguments without double-encoding;
• adds regression coverage for text, normalized structured output, and legacy tool-call responses.

🧪 Testing

• pytest tests/unit -q -o log_cli=false — 491 passed, 1 skipped
• focused LangChain utility tests — 3 passed
• Ruff 0.12.9 check and format — passed
• pre-commit on all changed files — passed
• real langchain_core AIMessage callback smoke test — passed
• git diff --check — passed

[gnanirahulnutakki]

Reviewed and reproduced on exact head 85f689612f3fff70274af24739bb22b522455418.

I verified the structured-output path with a real langchain_core AIMessage/ChatGeneration: normalized tool-call arguments are serialized once and emitted under the existing indexed gen_ai.completion.0.tool_calls.0.* conventions, while text completions retain role and finish-reason attributes.

Local verification:

• pytest tests/unit -q -o log_cli=false: 491 passed, 1 skipped
• focused LangChain utility tests: 3 passed
• Ruff 0.12.9 check and format check on all changed files: passed
• git diff --check: passed

No blocking issues found.