Demucs

**From this version, Python 3.7 is no longer supported. This is not a problem since the latest PyTorch 2.0.0 no longer support it either.**

Various improvements by CarlGao4. Support for `segment` param inside of HTDemucs
model.

Made diffq an optional dependency, with an error message if not installed.

Added output format flac (Free Lossless Audio Codec)

Will use CPU for complex numbers, when using MPS device (all other computations are performed by mps).

Optimize codes to save memory

Allow changing preset of MP3

Adding hybrid transformer Demucs model.

Added support for [Torchaudio implementation of HDemucs](https://pytorch.org/audio/main/tutorials/hybrid_demucs_tutorial.html), thanks skim0514.

Added experimental 6 sources model `htdemucs_6s` (`drums`, `bass`, `other`, `vocals`, `piano`, `guitar`).

Option to customize output path of stems (CarlGao4)

Fixed bug in pad1d leading to failure sometimes.

Added `--segment` flag to customize the segment length and use less memory (thanks CarlGao4).

Fix reflect padding bug on small inputs.

Added option to split into two stems (i.e. vocals, vs. non vocals), thanks to CarlGao4.

Added `--float32`, `--int24` and `--clip-mode` options to customize how output stems are saved.

Fix bug in weights used for different sources. Thanks keunwoochoi for the report and fix.

Improving drastically memory usage on GPU for long files. Thanks a lot famzah for providing this.

Adding multithread evaluation on CPU (`-j` option).