For speech-to-speech models like Gemini Live, audio goes directly to the
LLM and transcriptions arrive too late to be useful for turn decisions.
The existing TurnAnalyzerUserTurnStopStrategy waits for STT transcripts
before triggering end-of-turn, adding unnecessary latency.
This adds VADTurnAnalyzerUserTurnStopStrategy which triggers immediately
on turn analyzer COMPLETE without waiting for any STT transcript. Also
fixes the Gemini Live local VAD example to use UserStartedSpeakingFrame/
UserStoppedSpeakingFrame instead of VAD variants, since with local turn
management these are the frames that flow through the pipeline.
677ca04a18