WIP: [python-package] use sklearn_compat for multi-version scikit-learn compatibilities

[glemaitre]

This is a try to see if the tool provided in https://github.com/sklearn-compat/sklearn-compat could helped at maintenance.

For the moment this PR should be disregarded and do not need any review.

[jameslamb]

Thanks.

Before spending more time on this, please do remember what I said on Bluesky (link)... I personally would only support something like this if it looked liked it'd significantly reduce the maintenance burden of supporting the range of Python and scikit-learn versions we are trying to support. You're welcome to use our CI for testing this, and I'll respect your request to not review it until you feel it's ready.

And please, run the following and commit it to this branch to save some CI resources for other development in the project (these jobs should not be relevant for this work):

git rm .appveyor.yml .github/workflows/r_package.yml .github/workflows/cuda.yml

[glemaitre]

Before spending more time on this, please do remember what I said on Bluesky (link)... I personally would only support something like this if it looked liked it'd significantly reduce the maintenance burden of supporting the range of Python and scikit-learn versions we are trying to support. You're welcome to use our CI for testing this, and I'll respect your request to not review it until you feel it's ready.

@jameslamb I completely understand the trade-off and do not expect the PR to get through if it is not looking alright on your side. Up to now, I see that the main bottleneck was validate_data, the tags, and the tests. So at the end, you workaround could involve less changes that the file that we vendor.

I initially wanted to use the CI to assess how far we are with the compatibility file that we wrote so far.

And please, run the following and commit it to this branch to save some CI resources for other development in the project

Thanks for the pointers. I'll bypass those checks.