feat: add MOSS diarization backend

[LauraGPT]

Summary

• add opt-in --model moss support through the tested vLLM transcription API
• preserve MOSS speaker segments for SRT and speaker-based clipping without external VAD or speaker models
• reject truncated final segments instead of silently returning partial long-form transcripts
• document OpenMOSS ownership, pinned model revision, deployment boundary, and Paraformer fallback for precise arbitrary text clipping

This integrates the third-party OpenMOSS model without presenting it as a FunASR/FunClip-owned checkpoint. Follow-up context: OpenMOSS/MOSS-Transcribe-Diarize#48.

Validation

• python -m pytest -q tests: 84 passed, 1 skipped
• focused Ruff checks and compileall passed
• live H100 + vLLM inference on a two-speaker WAV returned S01/S02 segments, generated SRT, and clipped S02 to the expected 53,440 samples
• exact commit: 04b478225f67fd0892467d9810d26276ab25631d

Boundaries

• the packaged CLI exposes only the end-to-end tested vLLM path
• MOSS timestamps are segment-level; precise arbitrary text clipping remains on Paraformer
• no external vad_model or spk_model is attached, preserving global speaker identity

Signed-off-by: LauraGPT 18321252+LauraGPT@users.noreply.github.com