Backend.logsumexp: keepdims missing from Cupy and Tensorflow backends; Cupy reimplements cupyx.scipy.special.logsumexp

[tvercaut]

Describe the bug

Backend.logsumexp is declared as logsumexp(self, a, axis=None, keepdims=False) and its docstring states that it follows the scipy.special.logsumexp API. NumPy (ot/backend.py:1342), Jax (:1782) and Torch (:2460) honour that signature, but:

• CupyBackend.logsumexp (:2949) and TensorflowBackend.logsumexp (:3413) are declared (self, a, axis=None) and raise TypeError on keepdims=.
• CupyBackend.logsumexp is a hand-rolled port of a 2021 SciPy snippet. cupyx.scipy.special.logsumexp(a, axis=None, b=None, keepdims=False, return_sign=False) now exists and is a drop-in, matching what the NumPy backend does with scipy.special.logsumexp.
• tf.math.reduce_logsumexp already accepts keepdims=, so the TF gap is a one-word fix.

No current caller passes keepdims, so nothing is broken today — this is a latent API-conformance inconsistency, not a live bug. But it silently blocks backend-agnostic code from using keepdims, and test_backend.py only exercises nx.logsumexp(Mb) with no arguments, so the divergence is invisible to CI.

To Reproduce

1. Call ot.backend.TensorflowBackend().logsumexp(a, axis=0, keepdims=True) (or the equivalent on CupyBackend).
2. Observe TypeError: logsumexp() got an unexpected keyword argument 'keepdims'.

The same call succeeds on the NumPy, Jax and Torch backends.

Expected behavior

All backends should accept the same logsumexp(self, a, axis=None, keepdims=False) signature declared on the base Backend class, matching the scipy.special.logsumexp API the docstring already promises.

Environment

• Affects all environments; this is a static signature/API-conformance issue, not environment-specific.

Additional context

I plan to open a small PR fixing this (aligning Cupy on cupyx.scipy.special.logsumexp and adding keepdims to the Tensorflow backend), as a prerequisite for some N-D refactoring work in ot/bregman/_convolutional.py that needs a backend-agnostic logsumexp.