fix: improve Telugu accuracy and final transcript reliability

Summary: Improvements to Telugu live transcription pipeline to enhance reliability, stability, and accuracy in WebSocket streaming without changing the ASR model.

Changes:

• Improved WebSocket session handling:
  • clean new session initialization
  • reset support for new recordings
  • safer finalization before socket close
  • better logging for chunk processing and responses
• Improved transcription pipeline:
  • partials use fast PCM-based path for low latency
  • final output uses full transcribe_audio(...) path for better accuracy and punctuation
• Added streaming configuration:
  • partials toggle (STREAMING_ENABLE_PARTIALS)
  • increased partial and silence thresholds for more stable output
• Improved speech detection and buffering for cleaner final results

Impact:

• Fixes missing final transcript in UI after stop
• Reduces noisy/unstable partial outputs
• Improves streaming session reliability
• Enhances transcription accuracy without model change

Testing:

• Verified WebSocket flow and final transcript delivery
• Confirmed no premature socket closure
• Backend runs without errors after changes

Closes #15