Improvements of the checkpoint functionality and memory occupation of DER

[lrzpellegrini]

The main focus of this PR is to introduce various fixes and improvements for the checkpointing functionality:

• The checkpoint functionality now supports a simplified way to register dataset objects so that only constructor parameters are stored
  • This helper is now applied to more datasets
• A de-duplication utility has been added to prevent the in-memory duplication of dataset objects
• Greatly improve the performance of LazyIndices by switching to a NumPy-based storage of indices. This also greatly reduces the time taken to save and load checkpoints
  • Maximum depth before eagerifying reduced to 2
  • Added an additional unit test for LazyRange
• Loading a checkpoint in PyTorch 1.13.* is now handled by a fixed version of torch.load(...) (which is bugged in PyTorch 1.13.*)
• Re-introduced the device re-mapping functionality
• Checkpointing has been moved to a separate module to prevent import issues
  • Adapted RTD documentation

Additional elements:

• Fixes to DER memory occupation and checkpoint size:
  • FlatData can now discard (remove the reference to) unused elements (must be enabled via a specific constructor parameter)
    • Useful when storing large amounts of data as dataset data attributes (such as in DER, where logits were saved for each training exemplar)
  • Adapted DER strategy to use this functionality

Misc:

• Added AvalancheImageNet dataset, a clone of torchvision ImageNet dataset that supports an additional meta_root parameter. Useful when working on certain HPC clusters.

[AntonioCarta]

Thanks, all of these changes are great. Unfortunately, I messed up the CI by trying to add support for 3.11. Maybe let's wait until I fix it before merging this PR.

[coveralls]

Pull Request Test Coverage Report for Build 7653925810

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.4%) to 73.486%

---

Totals
Change from base Build 7653912528:	-0.4%
Covered Lines:	19049
Relevant Lines:	25922

---

💛 - Coveralls