fix: filter silence hallucinations from cloud transcription models

[SeoFood]

Summary

Fixes #19. Certain Whisper models (Large V3 Turbo, GPT-4o Transcribe) hallucinate random multi-language text when given silent audio. This adds two layers of silence detection:

• Energy gate (pre-filter): Tracks pre-gain peak RMS per audio chunk. If no chunk during the recording exceeds the speech energy threshold (0.01 RMS), transcription is skipped entirely and "No speech detected" is shown. This works correctly even when AGC/WhisperMode amplifies the audio, since it checks the raw microphone level.
• no_speech_prob filter (post-filter): Parses the no_speech_prob field from verbose_json Whisper API responses (OpenAI whisper-1, Groq, OpenAI Compatible). If all segments report > 0.8 non-speech probability, the result is discarded as a hallucination.

Both filters apply to the final transcription path (ProcessSingleJobAsync) and the live polling fallback (RunPollingFallbackAsync). WebSocket streaming providers (Deepgram, AssemblyAI) handle silence server-side and are unaffected.

Test plan

• Record 2-3 seconds of silence with Groq Whisper V3 Turbo - should show "No speech detected"
• Record silence with GPT-4o Transcribe - should show "No speech detected"
• Record normal speech - should transcribe correctly (no false positives)
• Verify whispered speech is still captured (low energy but above threshold)
• Verify local models (Parakeet, Canary) still work correctly via VAD path