refactor sinkhorn_divergence

[marcocuturi]

Ensure that sinkhorn_divergence function in tools returns a tuple. First element in Tuple (what most users will want) is divergence float value. Second element is detailed output.

This new convention will likely trigger

AttributeError: 'tuple' object has no attribute 'divergence'

errors in code that is using the former API, apologies for this!

Also, we may think about featuring the wrappers in progot.py L. 410 or appearing also in sinkhorn_divergence_test.py L.432 more prominently, as they follow the more intuitive API sdiv(x,y) which our implementation does not allow directly, as we need to pass geometry type.

As I believe 99% of users use Sinkhorn divergences with a point cloud geometry, we might want to remove that ambiguity/possibility?

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[codecov]

Codecov Report

Attention: Patch coverage is 66.66667% with 2 lines in your changes missing coverage. Please review.

Project coverage is 87.83%. Comparing base (27b639e) to head (e7b8bfd).
Report is 30 commits behind head on main.

Files with missing lines	Patch %	Lines
src/ott/tools/progot.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #577      +/-   ##
==========================================
- Coverage   87.84%   87.83%   -0.01%     
==========================================
  Files          73       73              
  Lines        7823     7826       +3     
  Branches     1127     1127              
==========================================
+ Hits         6872     6874       +2     
- Misses        798      799       +1     
  Partials      153      153              

Files with missing lines	Coverage Δ
src/ott/tools/sinkhorn_divergence.py	91.86% <100.00%> (+0.19%)	⬆️
src/ott/tools/progot.py	29.16% <0.00%> (-0.21%)	⬇️

[michalk8]

Would simplify to (also in the one other place), but we can also keep:
The Sinkhorn divergence and the output object.

[marcocuturi]

thanks.

i have also made a small change. _sinkhorn_divergence now only outputs the output object, it's sinkhorn_divergence that chooses to expose explicitly the value first in the Tuple, with the output second.