Fall back to librosa onset detection when madmom is unavailable

Add missing madmom dependency for onset mask

Fix ZeroGPU pickling error by extracting plain values before GPU boundary

Fix ZeroGPU pickling error and allowed_paths

Fix torch.load weights_only for PyTorch 2.6+

Loosen pydantic version to resolve gradio conflict

Add WhAM Gradio app with ZeroGPU support

initial commit