colab: subtitle-only defaults and whisperx robustness

colab: pin torch cuda121 and skip whisperx/demucs deps

Pin numpy below 2.1 in Colab setup

Force overwrite ct2 output dir

Force torch load non-weights-only in whisperx

asr: force torch.load weights_only off for ct2 convert

asr: download missing hf weights for ct2 conversion

asr: fix ct2 converter output dir

asr: auto-convert whisper models to ctranslate2

colab: install iputils-ping

Patch sitecustomize path in Colab

Ensure torchaudio backend common stubs always updated

Stub torchaudio AudioMetaData for WhisperX

Stub torchaudio.backend.common for Colab

Patch torchaudio backend module for Colab