Refactor TranscriptionUserTurnStopStrategy and TurnAnalyzerUserTurnStopStrategy
to use VADUserStoppedSpeakingFrame as the ground truth for when speech ended,
rather than triggering timeouts from transcription frames.
If we aggregate transcriptions we will get incorrect interruptions. For example,
if we have a strategy with min_words=3 and we say "One" and pause, then "Two"
and pause and then "Three", this would trigger the start of the turn when it
shouldn't. We should only look at the incoming transcription text and don't
aggregate it with the previous.
