Drop support for the redundant and deprecated cupy.array_api in favor of array_api_compat.

[ogrisel]

While reviewing #29615 I realized that we don't necessarily need to maintain our own sklearn.utils._array_api._clip fallback since we can instead rely on array-api-compat in recent versions.

It's still a bit verbose to use to correctly handle non-cpu devices with PyTorch but I think this should be fixed upstream:

• Automatically use the correct device in xp.clip with passed Python number literal as bounds data-apis/array-api-compat#177

EDIT: the scope of this PR has widened as explained in #29639 (comment).

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: d8a04d3. Link to the linter CI: here}

[ogrisel]

v1.6.rst already mentions cosine_distances for #29265. I don't think it's worth cluttering the changelog with 2 PRs for this function.

[ogrisel]

Actually, array_api_compat does not yet provide a xp.clip for the CuPy wrapper. So we better stick to using our own _clip fallback for the time being.

Here is the error reported by our CUDA CI on CuPy inputs:

AttributeError: module 'cupy.array_api' has no attribute 'clip'. Did you mean: 'flip'?

EDIT: this is unexpected because the changelog of array-api-compat 1.8 explicitly mentions the clip function.

EDIT 2: I also tried the following snippet after install array-api-compat 1.8 on google colab and it works as expected:

import array_api_compat.cupy as xp

data = xp.linspace(0, 1, num=5)
xp.clip(data, 0.1, 0.9)

Google Colab is running with CuPy 12.2.0 installed by default at the time of writing.

EDIT 3: I understand what's going on: the array_api_compat.cupy (for the cupy module) is up to date but the cupy.array_api module (that is not wrapped by array_api_compat is not yet updated to implement clip.

We could either add a clip method to our internal sklearn.utils._array_api._ArrayAPIWrapper class that is currently only used to handle missing stuff in cupy.array_api.

Or alternatively we could decide to no longer run the tests for cupy.array_api and instead recommend cupy users to use array_api_compat.cupy instead of cupy.array_api.

[ogrisel]

Or alternatively we could decide to no longer run the tests for cupy.array_api and instead recommend cupy users to use array_api_compat.cupy instead of cupy.array_api.

Since numpy.array_api was removed in numpy 2, I expect a similar fate for cupy.array_api. I opened cupy/cupy#8470 to check whether is this is planned or not.

[ogrisel]

Since we have the confirmation from upstream that the experimental cupy.array_api will be sunset in a near future release and since array_api_compat.cupy works well for inputs directly passed as cupy arrays, let's remove the code dedicated to testing and bringing compat wrapping for this module (see 360b943). Since array API support is experimental in scikit-learn, we don't need to follow the usual deprecation cycle for this removal.

This removes quite a bit of complexity / technical debt and should also speed up our test suite by not testing for cupy twice.

I will push an update to the changelog in a forthcoming commit.

[ogrisel]

Note to reviewers: cupy/cupy#8180 was fixed and will be included in the next release (v14) but we still have to wait for the time being.

[ogrisel]

I believe this is ready for review.

[thomasjpfan]

This LGTM. I'm happy that we can finally remove _ArrayAPIWrapper.

[OmarManzoor]

LGTM. Thank you @ogrisel