perf: int8/fp16 quantisations + configurable num_step (dub.py)

perf(dub): overlap original-video upload with separation; non-blocking per-segment uploads

fix(ingest): glob for yt-dlp's actual output filename (.mp4/.mkv/.webm)

fix(ingest): yt-dlp via sys.executable -m yt_dlp; clearer subprocess errors

Add /v1/dub/logs/tail endpoint for editor log drawer

dub: drop torchcodec stub __getattr__ raise (was breaking torch internals)

dub: stub torchcodec at sys.modules to bypass FFmpeg4/torch2.7 incompat

dub: add huggingface_hub>=1.0 use_auth_token shim for pyannote 3.3.x

add dub.py for self-bootstrap

add server.py for self-bootstrap

Upload OmniVoice weights and tokenizer snapshot

Duplicate from k2-fsa/OmniVoice