feat: use `htdemucs` model for stem separation instead of `htdemucs_ft`

Fix: use librosa.load instead of torchaudio.load (torchcodec can't open MP3)

Fix: remove invalid condition_on_previous_text kwarg; fix torch_dtype deprecation

Fix: defer CUDA usage to inside @spaces.GPU; use soundfile backend for torchaudio.load

Upload app.py