[MRG] Fix screenkhorn with the full budget on non-square problems

[itzzdev09]

Types of changes

• Bug fix

Description

ot.bregman.screenkhorn fails when ns_budget and nt_budget are the full sizes and the problem is not square:

a, b = ot.unif(5), ot.unif(4)
M = ot.dist(np.random.randn(5, 2), np.random.randn(4, 2))
ot.bregman.screenkhorn(a, b, M, 1.0, ns_budget=5, nt_budget=4)
# ValueError: shapes (5,) and (4,) not aligned: 5 (dim 0) != 4 (dim 0)

In the full-budget branch, vec_eps_IJc and vec_eps_IcJ were created with each other's size. In the general branch, vec_eps_IJc sums K_IJc over its columns, so it has one entry per kept source point, and vec_eps_IcJ has one per kept target point. The full-budget branch used nt and ns the other way round, which only worked when ns == nt. With the sizes swapped back, the full-budget result matches ot.sinkhorn.

How has this been tested

Added test_screenkhorn_full_budget_non_square (20 x 12, full budgets), which compares with ot.sinkhorn. It fails on master, and the existing screenkhorn test still passes.

Checklist

• I have added tests to cover my changes.
• I have updated the RELEASES.md file.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.92%. Comparing base (98d09a1) to head (d8c942c).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #871      +/-   ##
==========================================
+ Coverage   96.86%   96.92%   +0.05%     
==========================================
  Files         128      128              
  Lines       26304    26312       +8     
==========================================
+ Hits        25480    25503      +23     
+ Misses        824      809      -15     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.