Perturbation space - pseudobulk, discriminative classifier, clustering

[AlejandroTL]

PR Checklist

Description of changes

Update of the PerturbationalSpace, including:

• Pseudobulk
• Discriminative classifier (includes Pytorch Lightning models and trainer)
• Clustering K-Means and DBSCAN

[Zethson]

Thank you very much!

Things I did:

1. Solved all pre-commit errors. You seem to like mutable defaults (e.g. populated lists) for argument defaults. Don't do this, please :)
2. Moved all Made the PerturbationSpace class abstract and all other classes are now implementing it and will be called standalone as discussed
3. Moved the _utils stuff into the discriminator file because this is where it's supposed to live
4. Random minor stuff

General comments:

1. Don't run any functions explicitly in test files. pytest picks them up and runs them on its own.
2. The tests are currently bad for two reasons: 1. Some of them are broken and don't even run; 2. They should actually test the calculated perturbation space. Currently you're just running the function and then hoping for the best. Create artificial data that allows you to test for whether the calculated perturbation space makes sense. e.g. 3 very distinct clusters of cells and then measure the R2 between the 3 points of the perturbation space that you calculated and it should be at least SOMENUMBER or so.
3. Ensure that every function signature that people will use (call) primarily is similar. All of them should have a copy parameter, a layer parameter, .... you get the gist. The experience when working with the spaces should be very consistent.
4. Document the functions properly with docstrings. We'll be exposing some functions that are not yet documented (all call primarily)
5. Do NOT add type hints in docstrings (as shitty VScode does automatically). We have them in the function signature and Sphinx will renders this eventually.

Please resolve my comments as you go through them. In an ideal world (that we 100% do not have due to lack of time) one would add one commit per comment.

If stuff is unclear - ask away!