RAPIDS 26.02 support

[jameslamb]

Description

Running xgboost 3.1.3 tests against RAPIDS 26.02, I found 2 issues that I think are still present on master:

• cudf removed the mode parameter to .to_pylibcudf()
  • PR: Remove exposure on column construction and unwrap buffers on pylibcudf conversion rapidsai/cudf#20980
  • already reported in xgboost 3.2.0 crashes with cudf 26.02 when there are categorical features #12138 and fix up in Fix categorical data support with cuDF 2602. #12140
• cudf removed from_pandas() methods, in favor of passing pandas Dataframes to constructors like cudf.DataFrame()
  • PR: Remove deprecated .from_pandas constructors rapidsai/cudf#20925

Opening this to report those and to have a place for you all to track full RAPIDS 26.02 support.

Notes

Looks like this PR would add CI images with RAPIDS 26.02 that'd help with testing: dmlc/xgboost-devops#78

For the .from_pandas() changes, stuff like this:

xgboost/python-package/xgboost/testing/updater.py

Line 365 in 6e3b9db

X = cudf.from_pandas(X)

Should be converted to like

 X = cudf.DataFrame(X) 

Though I'm not sure how many cuDF releases back that pattern is supported, so don't know if you can make that change unconditionally. @mroeschke would know.

[mroeschke]

cudf removed from_pandas() methods, in favor of passing pandas Dataframes to constructors like cudf.DataFrame()

I opened #12144 to fix this usage (note that the cudf.from_pandas function remains but the classmethod on objects like cudf.DataFrame.from_pandas was removed)