[MRG] Support integer class labels in CUDA domain adaptation

[antonsoo]

Types of changes

Bug fix with regression tests and a release note.

Motivation and context / Related issue

Fitting a domain-adaptation estimator with CUDA integer labels fails while
forming the label-presence mask:

import torch
import ot

X = torch.tensor([[0.0], [0.1]], dtype=torch.float64, device="cuda")
y = torch.tensor([0, 1], device="cuda")
transport = ot.da.EMDTransport().fit(Xs=X, ys=y, Xt=X, yt=y)
# Before: RuntimeError: "addmm_cuda" not implemented for 'Long'
# After: fit completes and coupling_ is on CUDA.

The same failure occurs with int32 labels. Replace the (ns, 1) @ (1, nt)
matrix product with a broadcast product. Both compute the same outer product;
broadcast multiplication supports these integer types without changing dtype
or device. Missing-label behavior stays the same.

How has this been tested (if it applies)

• Four new CUDA cases fail on master and pass here: int32/int64 labels, with
  and without missing source/target labels. Four CPU controls pass throughout.
• python -m pytest test/test_da.py -q: 59 passed, 2 skipped.
• pre-commit run --all-files: passes.
• python -m pytest --durations=20 -q test/ --doctest-modules --maxfail=3: 2,705 passed, 62 skipped, 6 xfailed (NumPy and PyTorch backends, 235 seconds).

On actual Iris, Wine and handwritten-digit records, all 48 integer CUDA
EMD/Sinkhorn workflows now complete. Transport, barycentric mapping, propagated
label probabilities and downstream classifier predictions exactly match the
floating-label CUDA controls. The 48 CPU integer control records and 48 CUDA
floating control records are unchanged by the patch. Independently formed costs
and SciPy assignment objectives agree within 2.89e-15 and 2.78e-17 respectively.
This checks execution and equivalence, not classification accuracy. Validated on
Linux with torch 2.7.0+cu128 and an RTX 5070.

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made.
• All locally available tests passed, and additional code has been covered with new tests. See the exact backend scope and skips above.
• I have added the fix to RELEASES.md.