Torchcache

Added
- Use the `mmap` of Python to allow using mmapped loading for all versions of PyTorch, even with compression.

Changed
- The parameter documentation is moved from the underlying class to the exposed decorator, so that the signature hints can function properly.

Fixed
- API documentation not visible.

**Full Changelog**: https://github.com/meakbiyik/torchcache/compare/v0.3.2...v0.4.0

Fixed
- Performance regression due to `weights_only` in `torch.load`
- Allowed changing embedding shape on-the-fly

**Full Changelog**: https://github.com/meakbiyik/torchcache/compare/v0.3.1...v0.3.2

Empty release to fix publish workflow, see v0.3.0 release for the updates: https://github.com/meakbiyik/torchcache/releases/tag/v0.3.0

Added
- use mmap in `torch.load` if available (currently in nightly: https://github.com/pytorch/pytorch/pull/102549)
- finer controls for module cache: ability to set it manually it to prevent cache invalidation by mistake

Changed
- `ztsd` compression instead of brotli for faster decompression

Fixed
- Invalidate cache if some crucial torchcache requirements change (subsample count and compression)

**Full Changelog**: https://github.com/meakbiyik/torchcache/compare/v0.2.0...v0.3.0

**Full Changelog**: https://github.com/meakbiyik/torchcache/compare/v0.1.0...v0.2.0

First release 🚀