HTDemucs v4 — MLX conversion

Upstream model by Simon Rouard, Francisco Massa and Alexandre Défossez.
Copyright (c) Meta Platforms, Inc. and affiliates

Licensed under the MIT License (MIT).
Licence text: LICENSE (also at https://github.com/facebookresearch/demucs/blob/main/LICENSE)
Upstream: https://github.com/facebookresearch/demucs

Modifications
-------------
These weights are a modified version of Meta's HTDemucs v4 checkpoint,
converted to the MLX array format for inference on Apple silicon. Specifically:
the upstream pickled Torch .th checkpoint is read for its tensor state dict
only, weight keys are remapped to match this project's module names,
convolution weights are transposed where MLX's layout differs from PyTorch's
(both the 1-D and 2-D cases), fused QKV projections are split into separate Q,
K and V tensors, and the result is re-serialised as safetensors. The model is
re-implemented against MLX rather than PyTorch.

No weights are retrained or fine-tuned.

Tensor precision is unchanged: the upstream checkpoint stores float32 and so
does this conversion. No weight value is altered by rounding — the Demucs
converter performs no dtype cast.

This conversion is distributed by the mlx-audiogen project
(https://github.com/jasonvassallo/mlx-audiogen), which is itself Apache-2.0.
That licence covers the conversion code only — these weights remain governed by
the MIT License above.

The checkpoint this was converted from is distributed by Meta at
https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/955717e8-8726e21a.th,
referenced by the upstream project above.