fix(external): ensure solo soft predict returns probs

[martinkim0]

BREAKING CHANGE: SOLO.predict used to incorrectly return logits instead of probabilities when soft=True (the default option). this pr fixes that but it is a breaking change.

closes #2688

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 88.85%. Comparing base (1fff5b1) to head (b3db342).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2689      +/-   ##
==========================================
- Coverage   88.86%   88.85%   -0.02%     
==========================================
  Files         158      158              
  Lines       13247    13247              
==========================================
- Hits        11772    11770       -2     
- Misses       1475     1477       +2     

Files	Coverage Δ
scvi/external/solo/_model.py	96.02% <100.00%> (+0.66%)	⬆️

... and 1 file with indirect coverage changes

[canergen]

It needs to be labeled as breaking change. It might be some folks use softmax after SOLO (I do). I guess it's better to also print a warning about the change for scVI-tools 1.3.

[martinkim0]

It needs to be labeled as breaking change. It might be some folks use softmax after SOLO (I do). I guess it's better to also print a warning about the change for scVI-tools 1.3.

Right - I forgot to do that. This is slated to be released with 1.1.3, should we revert it then? It's technically a bug fix so IMO it should be fine.

[canergen]

I think it's fine to release with 1.1.3 and I prefer outputting probabilities. I agree the code was wrong and not the documentation. However, we should highlight this change as it might affect analysis scripts. I'm open whether it outputs a warning or not but the release note should contain a breaking change.

[martinkim0]

Makes sense! I'll add a warning and a note to the changelog.