fix(google): preserve non-blocking realtime generation

[usluio]

Summary

• keep Gemini realtime output streams open when a NON_BLOCKING function call arrives
• continue using the existing immediate finalization behavior for blocking function calls
• add regression coverage for both blocking and non-blocking tool behavior

Why

Gemini 2.5 asynchronous function calling allows the model to continue the active conversation while a non-blocking tool executes. The Google realtime plugin currently calls _mark_current_generation_done() for every tool call, which closes the audio and text streams before trailing output in the same turn can be delivered.

This change only defers finalization for declarations configured with Behavior.NON_BLOCKING. The generation still closes on Gemini's existing generation_complete and turn_complete events. Blocking calls retain the current behavior.

This is intentionally narrower than #6285, which addresses blocking/tool-terminated turns such as Gemini 3.1 with an audio-drain timeout.

Related: #5742

Tests

• uv run --python 3.12 pytest tests/test_plugin_google_realtime.py --unit -q
• uv run --python 3.12 ruff check livekit-plugins/livekit-plugins-google/livekit/plugins/google/realtime/realtime_api.py tests/test_plugin_google_realtime.py
• uv run --python 3.12 ruff format --check livekit-plugins/livekit-plugins-google/livekit/plugins/google/realtime/realtime_api.py tests/test_plugin_google_realtime.py
• uv run --python 3.12 mypy --untyped-calls-exclude=smithy_aws_core -p livekit.plugins.google

Full repository type checking is currently blocked locally by a missing boto3 stub in the AWS experimental plugin.

[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 1 additional finding.