Ignore cudf's __dataframe__ deprecation with simpler filters.

[bdice]

This is a follow-up to #6229. That PR should unblock CI, but reviewers would like simpler filters.

See:

• Ignore cudf's __dataframe__ deprecation. #6229 (comment)
• Ignore cudf's __dataframe__ deprecation. #6229 (comment)

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[betatim]

I updated this PR with the changes from #6227 which will, I think, fix the CI failures.

[bdice]

I removed a deprecation filter for a warning @csadorf and @jcrist said they could not reproduce. It occurred again in CI here. It happens in an optional job, perhaps there is a difference in the environment/dependencies/parameters/etc.

FAILED test_kneighbors_classifier.py::test_predict_non_gaussian[100-4-40-100] - FutureWarning: Support for loading dataframes via the `__dataframe__` interchange protocol is deprecated

https://github.com/rapidsai/cuml/actions/runs/12989973215/job/36226273552?pr=6239#step:8:5762

[betatim]

From the name of the CI job it sounds like it runs with cudf.pandas enabled, that might explain why locally they can't reproduce this. It looks like creating the cudf dataframes from cudf.pandas dataframes is what triggers the deprecation warning: X_device_train = cudf.DataFrame.from_pandas(X_host_train). Without the accelerator X_host_train is a pandas df, but with the accelerator it is a cudf.pandas df. My initial thought is that cudf.DataFrame.from_pandas needs updating to not use __dataframe__.

There is also a question for cudf.pandas dataframes: they should do what pandas does, either have a __dataframe__ without deprecation or not have it. At least it seems against its aim to behave differently from pandas.