[python] deprecate PyTorch loader prototype

[mlin]

The PyTorch loader built-in to the cellxgene-census API is "graduating" as TileDB-SOMA-ML. This adds runtime+docs deprecation notices (leaving it functional for now), and revises the tutorial notebook to use TileDB-SOMA-ML.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 91.94%. Comparing base (5817276) to head (95fc2b5).
Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1365   +/-   ##
=======================================
  Coverage   91.94%   91.94%           
=======================================
  Files          83       83           
  Lines        6742     6745    +3     
=======================================
+ Hits         6199     6202    +3     
  Misses        543      543           

Flag	Coverage Δ
unittests	91.94% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[ivirshup]

Just a minor point, @ebezzi will get back to you for a review on the notebooks

[ebezzi]

LGTM. The notebook seems clear enough. Just two points:

1. Would it be worth specifying that this notebook originally used another module that is now deprecated? I'm thinking that there might be some users who might be confused if they were using this as a reference for some of their work (we will announce the deprecation, but not many people follow our announcements).
2. Curious why the model trained with this dataloader seems to perform slightly worse than the previous one. Could be because this is a newer Census version? That said, I don't think this is worth an investigation. It would be nice if we could some kind of regression testing against a real model (e.g. scvi), but this shouldn't be a blocker.

[mlin]

@ebezzi thanks!

1. Good suggestion, added that note.
2. Yes, when I ran the notebook on the older Census version, the results were similar. The newer version has more cell types (and cells) answering the query, so in some sense the "baseline" difficulty of the problem is harder.