fix: ensure pipecat metadata is JSON

[duncankmckinnon]

Fix metadata serialization for non-dict LLM settings in pipecat instrumentor

Problem

When using pipecat with Google Gemini (or other providers with Pydantic model settings), the metadata span attribute was being set to a Python repr string like:

GoogleLLMSettings(model='gemini-3-flash-preview', extra={}, system_instruction=None, ...)

This caused export failures with StatusCode.UNKNOWN because the Arize backend requires metadata to be valid JSON.

Fix

In _attributes.py, the LLMServiceAttributeExtractor.extract_from_service method now converts non-dict settings objects to dicts before JSON serialization, using model_dump() (Pydantic v2), dict() (Pydantic v1), or __dict__ as fallbacks.

Testing

All 66 existing tests pass.

[github-actions]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.