[MRG] Use ot.utils.check_marginal for default marginals across solvers

[tomMoral]

Types of changes

• Refactoring / consistency (non-breaking, no new user-facing feature)

Motivation and context / Related issue

Follow-up to #851, which introduced ot.utils.check_marginal (fill a uniform
marginal on the inputs' backend/device when None, otherwise validate shape +
dtype/device) and used it in the batch solvers.

The same if a is None: a = unif(...) / nx.ones(...) / n idiom is repeated ~25×
across the rest of the library. Routing those through check_marginal makes
default-marginal handling consistent and adds early shape/dtype/device
validation for user-provided marginals (previously a mismatch surfaced as a deep
error further down the solver).

Description

Replaced the default-marginal idiom with check_marginal at 25 backend-preserving
1D sites:

• ot/gromov/: _gw.py, _partial.py, _semirelaxed.py, _lowrank.py
• ot/lowrank.py, ot/factored.py, ot/stochastic.py
• ot/bregman/_geomloss.py, ot/bregman/_barycenter.py
• ot/lp/_barycenter_solvers.py, ot/utils.py (split_sample_ratio)

check_marginal stays in ot.utils (not exported at top-level).

Deliberately not changed:

• ot/gaussian.py — weights are (n, 1) column vectors; unif/check_marginal
  normalize over the last axis, which would be wrong for that shape.
• ot/gromov/_dictionary.py, ot/weak.py, ot/bregman/_empirical.py — numpy-by-design
  paths (they convert everything to numpy first), so a CPU default is intentional.

One small hardening to check_marginal: when validating a provided marginal it
now infers the backend with get_backend(type_as, a), so a cross-backend
marginal raises the standard "Backends should be the same" error instead of an
AttributeError.

Note on behavior

These sites already passed type_as, so this is a consistency + validation
change, not a device bugfix. The one behavior change is that a provided marginal
whose shape/dtype/device disagrees with the reference now raises early. For the
common a=None path and for consistent-backend inputs (all existing tests) the
behavior is unchanged.

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check RELEASES.md).
• All tests passed, and additional code has been covered with new tests (helper already covered in [MRG] Fix device placement in batch Sinkhorn (solve_batch on GPU) #851).
• I have added the PR and Issue fix to the RELEASES.md file.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.85%. Comparing base (caa11a4) to head (60c1704).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #856   +/-   ##
=======================================
  Coverage   96.84%   96.85%           
=======================================
  Files         128      128           
  Lines       26195    26161   -34     
=======================================
- Hits        25369    25337   -32     
+ Misses        826      824    -2     

🚀 New features to boost your workflow:

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

[rflamary]

Thanks very much @tomMoral