ENH add shuffle to GroupKFold

[zvealey]

Reference Issues/PRs

closes #13619
Partially addressing #20520

What does this implement/fix? Explain your changes.

This update introduces a shuffle feature to GroupKFold, along with a new class, RepeatedGroupKFold, which supports repeating GroupKFold n times. When shuffle=False, the behavior remains unchanged, ensuring that groups are split to achieve folds of as equal size as possible. Setting shuffle=True shuffles the unique groups before they are assigned to folds, without prioritizing the size of the individual groups.

Any other comments?

Although enabling shuffling does not guarantee that folds will be of similar size, in scenarios where groups are roughly equal in size (a common situation) the size difference between folds should be minimal.

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: 3fa6a23. Link to the linter CI: here}

[NegatedObjectIdentity]

I would like to see such functionality in scikit-learn, since right now I use a external function to perform exactly this.

[glemaitre]

I see that I forgot to post my review that I started a while ago. @zvealey would you be able to address the comments?

[zvealey]

Hi @glemaitre! I missed when you posted your review! I updated the PR based on your comments!

[glemaitre]

I pushed a small commit just to ensure that one of the test was stable across a large amount of random seeds.

I took the opportunity to address a couple of comments that I would have raised.

[glemaitre]

LGTM. We will need a second review.

[glemaitre]

I pushed a small fix for the docstring.

[avm19]

Currently, as a result of this PR, the docstring says:

The folds are approximately balanced in the sense that the number of
samples is approximately the same in each test fold when shuffle is True.

but it should probably say "...when shuffle is False", because when shuffle is True the group sizes are not even computed in the code.

I guess, the typo comes from that it is easy to confuse balancing the number of groups and the number of samples. And I also got it wrong when writing a now-deleted comment and linking to the discussion above.

The best clarification is found in cross_validation.rst docs, and I suggest to copy this explanation verbatim into GroupKFold's docstring to avoid any confusion:

scikit-learn/doc/modules/cross_validation.rst

Lines 670 to 672 in 3fa6a23

	While :class:`GroupKFold` attempts to place the same number of samples in each
	fold when ``shuffle=False``, when ``shuffle=True`` it attempts to place equal
	number of distinct groups in each fold (but doesn not account for group sizes).