Initial work toward PyTorch data loaders

[bkmartinjr]

Notes to reviewers:

• APIs have changed from the CZI contribution. The demo notebook (tutorial_pytorch.ipynb) has been updated to run with all API modifications.
• This PR is for an initial drop we plan to use for pre-release feedback. More work is required prior to release
• This is set up as a separate Python package, with its own config/build/etc.

---

This PR contains a PyTorch iterable-style DataSet/DataPipe for use with SOMA Experiment. Initial code contributed by the Chan Zuckerberg single-cell team. Modifications to the original code include:

• stand-alone Python package tiledbsoma_ml
• significant refactoring to improve performance
• multi-worker/multi-GPU support - should work with DDP, DataLoader and Lightning
• CI and other config for stand-alone Python package
• improved context/config handling - pass through all user-specified configuration allowing use with other storage solutions, AWS multi-region, etc.
• fix lint and typing issues identified by the CI pipeline in this repo
• update doctrings and copyright
• some minor enhancements to unit tests
• added torch.utils.data.IterableDataset alongside the existing torchdata.datapipes.iter.IterDataPipe, acknowledging that the torchdata.datapipes is deprecated and slated for removal
• add Python 3.8 support
• API refinement to better align with ExperimentAxisQuey
• fix RNG state handling which caused incorrect results when shuffle=True

API change summary:

• constructor takes ExperimentAxisQuery, rather than than an Experiment
• reworked I/O and shuffle chunk size param for better API UX
• added "set_epoch" endpoint to allow distributed shuffle