Add recording health monitoring and teardown safeguards

[Gokias]

Motivation

• Improve robustness of live transcription by detecting when a connected VoiceClient has stopped recording and avoid silent failures.
• Give short transient gaps a configurable grace period before escalating to session teardown.
• Provide diagnosable logs and ensure session teardown fully cancels background tasks and removes state.

Description

• Add TRANSCRIBE_RECORDING_FAILURE_GRACE_INTERVALS config to tolerate short blips before escalation.
• In transcription_live_loop, when vc is connected but vc.recording is false increment session.recording_failure_count, emit a warning including guild_id, voice_state, recording_failure_count, and queue_depth, and apply the grace-period retry logic before proceeding to teardown.
• When session.recording_failure_count exceeds TRANSCRIBE_MAX_FAILURES, call teardown_transcription_session_for_recording_failure(...) with a clear failure reason containing consecutive failure count and queue depth.
• Reset session.recording_failure_count to 0 and log a recovery event whenever recording resumes/capture succeeds.
• Harden teardown_transcription_session_for_recording_failure to cancel and await loop_task and worker_task (avoiding cancelling the current task), then continue disconnect/notification and remove the session via remove_transcription_session.

Testing

• Performed static syntax check with python -m py_compile poopbot.py, which completed without errors.

---

Codex Task