fix(google_gemini): fix non-streaming response persistence and add usage tracking

[olivier-lacroix]

This PR fixes the issue where model responses would disappear from the browser chat window when streaming was disabled.

Problem:

In non-streaming mode, returning a dictionary (OpenAI-compatible format) was leading to empty content fields in subsequent background events (like chat:outlet or follow-up question generation). When these events reached the browser, they would overwrite the previously displayed response with an empty string, causing it to vanish.

Solution:

1. Return a String: Switched the return type of the pipe method to a simple string for non-streaming calls. Open WebUI recognizes this as a finalized message and definitively sets the message content, preventing background tasks from overwriting it.
2. Usage Tracking: Since we are now returning a string, we manually emit the usage data via the __event_emitter__ right before the return. This ensures token counts are still captured and saved to the database.
3. Avoid Duplication: Removed manual chat:message event emission in the non-streaming path as the string return already handles UI display correctly.