fix(google): capture usage_metadata before early continues in streaming

[Panmax]

Summary

• Fix prompt_tokens: 0, completion_tokens: 0 in LLM metrics for Gemini streaming requests
• Move the usage_metadata check before the two continue statements in LLMStream._run so token counts are always captured
• In Gemini's streaming API, the final chunk typically carries usage_metadata with complete token counts but has empty candidates or content.parts, causing the existing code to skip the usage check entirely

Details

In livekit-plugins-google/livekit/plugins/google/llm.py, the streaming loop has two early continue statements:

1. if not response.candidates: continue
2. if not candidate.content or not candidate.content.parts: continue

Both skip past the usage_metadata check at the end of the loop body. Since Gemini's final streaming chunk often has a finish_reason but empty content, the usage data (which is only fully populated on this last chunk) is never captured.

The fix moves the usage_metadata check to immediately after the prompt_feedback check, before any continue statements.

Test plan

• Existing tests pass
• Lint and type checks pass
• Manual verification: run a Gemini streaming session and confirm LLM metrics show non-zero prompt_tokens, completion_tokens, and tokens_per_second

🤖 Generated with Claude Code

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.