- Properly extract finish_reason from candidate object and map to OpenAI format
- Correctly extract usage metadata from response.usage_metadata structure
- Extract prompt_token_count, candidates_token_count, and total_token_count
- Add logging for usage metadata extraction
- Handle Google finish reasons: STOP, MAX_TOKENS, SAFETY, RECITATION, OTHER

This fixes the issue where Gemini responses were arriving corrupted to
OpenAI-compatible clients due to incorrect parsing of the new Google GenAI
SDK response structure.