Add FunASR/SenseVoice API backend

[LauraGPT]

Summary

• add a first-class FunASR API model backed by a separately running FunASR/SenseVoice OpenAI-compatible server
• support file/folder imports, live recording, and the funasrapi CLI model type
• add configurable base URL and model settings, with environment overrides for protected or remote endpoints
• keep FunASR credentials isolated from the saved OpenAI key and enforce the backend's transcription-only contract
• document server setup and usage across preferences, file import, live recording, and CLI guides

Boundary fixes

• allow live recording to start without a local model path for FUNASR_API, matching the existing OpenAI API behavior
• disable unsupported word-level parsing for FunASR responses so a text-only or empty-segment response cannot raise TypeError or discard the transcript
• omit prompt and timestamp_granularities fields that the FunASR endpoint does not support

Verification

• TDD live-recording regression failed before the fix because RecordingTranscriber was called 0 times; it now starts once with the expected API options
• TDD word-timing regression reproduced TypeError: 'NoneType' object is not iterable; it now returns the complete text fallback and sends no word-timing request
• focused FunASR boundary suite: 13 passed
• complete recording-widget module: 96 passed
• broad repository suite: 1037 passed, 4 skipped
• Ruff checks on both changed production modules
• git diff --check

The current GitHub Actions and Snapcraft runs are action_required with zero jobs, so they still require maintainer approval before GitHub can execute them.

Closes #1140

[LauraGPT]

Follow-up for head c7951ec: review found and fixed two contributor-side boundaries before maintainer review:

• live recording now allows the empty local model path required by FUNASR_API;
• FunASR file transcription disables unsupported word-level parsing and preserves text-only/empty-segment responses instead of raising TypeError or returning no transcript.

Current local verification is green: 13 focused FunASR tests, 96 recording-widget tests, and the broad repository suite (1037 passed, 4 skipped), plus Ruff on the changed production modules and git diff --check.

The new CI and Snapcraft runs are both action_required with zero jobs, so they need maintainer approval before GitHub can execute them. No rush given your summer schedule; approval and review whenever convenient would be appreciated.

[raivisdejus]

Summer is busy time, but sure, I will review this.

[LauraGPT]

Refreshed this PR onto current main and resolved the only merge conflict in buzz/transcriber/recording_transcriber.py.

Conflict resolution kept the upstream use_transcriptions_endpoint guard for local whisper.cpp and preserved the FunASR API behavior that routes translate requests back to /audio/transcriptions when the configured API does not support translation.

Validation on exact head f1abc011c561:

• python3 -m py_compile buzz/transcriber/recording_transcriber.py buzz/transcriber/audio_api.py
• conflict-marker check on buzz/transcriber/recording_transcriber.py
• QT_QPA_PLATFORM=offscreen PYTHONPATH=demucs_repo:deepmultilingualpunctuation:ctc_forced_aligner:whisper_diarization:$PWD /cpfs_speech/user/zhifu.gzf/codebase/Buzz-funasr-audit-20260714/.venv/bin/python -m pytest tests/transcriber/audio_api_test.py tests/transcriber/recording_transcriber_test.py tests/recording_transcriber_test.py -q -> 61 passed
• git diff --check

GitHub now reports the PR MERGEABLE; there are no configured CI checks on the PR.