feat(model): support batch key in condscvi

[martinkim0]

closes #2623

[martinkim0]

@canergen The main backwards compatibility concern right now is that, with the current AnnDataManager setup with a batch key, the model will receive a tensor of zeroes for batch_index even if batch_key is None.

There's not a straightforward way to distinguish this case from the case where a batch_key is registered but all observations have the same batch assignment. However, we'd like to treat the latter differently from the former because of backwards compatibility.

This is because a batch_index of zeros will still be passed into the encoder (if encode_covariates is True) and the decoder, and will affect model training. So anyone repeating model training from a previous version will see different results.

There are two potential ways to address this:

1. Modify CategoricalObsField so that it is evaluated to None in AnnDataLoader (similar to CategoricalJointObsField) if registry_key is None. More complicated but the better long-term solution. I need to look into this one more.
2. Just don't register a new CategoricalObsField if batch_key is None in CondSCVI.setup_anndata. Easiest solution but you have to do a couple of checks in CondSCVI and VAEC, potentially buggy.

[codecov]

Codecov Report

Attention: Patch coverage is 97.56098% with 1 lines in your changes are missing coverage. Please review.

Project coverage is 88.68%. Comparing base (6bdf1ec) to head (0d5a75d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2626      +/-   ##
==========================================
- Coverage   89.18%   88.68%   -0.51%     
==========================================
  Files         159      159              
  Lines       13308    13306       -2     
==========================================
- Hits        11869    11800      -69     
- Misses       1439     1506      +67     

Files	Coverage Δ
scvi/model/_condscvi.py	96.73% <100.00%> (ø)
scvi/module/_vaec.py	84.70% <97.14%> (-1.51%)	⬇️

... and 1 file with indirect coverage changes

[martinkim0]

@canergen Here's an initial exploration with using the batch key. TLDR Seems fine but can lead to overintegration when training for too long.
condscvi_batch.pdf

[martinkim0]

Losses are very similar across the runs:

[canergen]

We have quite some different datasets to see better integration. You need to run one UMAP per coarse celltype above. I don't think those are necessary but I requested only support to get the key into DestVI :).