[WIP] add mass feature to nx.kl_div and harmonize kl computation in the toolbox

[cedricvincentcuaz]

Types of changes

• Add feature mass=False|True in nx.kl_div to compute the generalized kl divergence either with the difference of total masses or not.
• Screened the toolbox to ensure proper use of nx.kl_div and along the way i set all existing offsets to nx.log to 1e-16.

Motivation and context / Related issue

How has this been tested (if it applies)

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR and Issue fix to the RELEASES.md file.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 96.67%. Comparing base (e530985) to head (368f979).

Additional details and impacted files

@@ Coverage Diff @@ ## master #654 +/- ## ======================================= Coverage 96.66% 96.67% ======================================= Files 85 85 Lines 16934 16956 +22 ======================================= + Hits 16370 16392 +22  Misses 564 564 

[mbuze]

Hi, this is a fantastic addition! I am currently trying to use sinkhorn_unbalanced2 and would love to have mass=True turned on for the marginal penalisation, but I am struggling to get it to work. Any tips on how to achieve it without much hacking?

(Sorry if I am being confused here -- potentially this does not affect the optimisation routine to find the optimal plan and so it really is just about returning the correct UOT cost?)