fix(rtzr): finalize utterances without reconnecting

[kimdwkimdw]

Problem

RTZR's flush() sends EOS, closing the WebSocket at each utterance boundary and requiring a new connection for the next utterance. RTZR supports {"type":"Finalize"} to complete an utterance while keeping the connection open.

Receive failures are also logged without reaching the recognition stream, so transcription can stop without triggering LiveKit's error handling.

Fix

Send buffered audio followed by Finalize on flush(). Reuse the connection across utterances and send EOS when input ends or after 25 seconds without audio. Graceful shutdown drains final responses; aclose() cancels immediately.

• Propagate receive and shutdown failures, preserve pending input across initial connection retries, and report final-response failures after input exhaustion as terminal errors.
• Reuse the shared progressive audio buffer, validate mono input before resampling, and count sent audio usage once.
• Refresh tokens before expiry, retry WebSocket authentication once on 401, honor endpoint overrides, and correct per-stream language and word timing metadata.

Replaces #4940.

References

• https://developers.rtzr.ai/docs/en/stt-streaming/#ai-voice-agent