ENH Add Array API compatibility for additive_chi2_kernel

[Tialo]

Towards #26024

Reference Issues/PRs

What does this implement/fix? Explain your changes.

Any other comments?

Due to _chi2_kernel_fast written in cython it can't be used with array-api containers. So I implemented all calculations using array broadcasting. It is not faster than cython function, so I use _chi2_kernel_fast when additive_chi2_kernel gets a numpy array.

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: edef1c9. Link to the linter CI: here}

[ogrisel]

LGTM. Could you please just update the _array_api.rst documentation file to reference this function?

[ogrisel]

And there are a bunch of conflicts to resolve when merging main.

[ogrisel]

Also please check that the CUDA tests pass with https://gist.github.com/EdAbati/ff3bdc06bafeb92452b3740686cc8d7c.

[Tialo]

CUDA tests passed

[ogrisel]

Ping @betatim @OmarManzoor.

[betatim]

I resolved the merge conflict and started the CUDA CI https://github.com/scikit-learn/scikit-learn/actions/runs/9501959732

edit: CI failed because I made a mistake when resolving the conflict.

[OmarManzoor]

@ogrisel @betatim A test seemed to be failing i.e. the one using Torch mps. I have proposed a fix here #29256. Could you kindly confirm that the error occurs and the fix is valid?