use CUDA-suffixed XGBoost wheels (#7550)

Contributes to rapidsai/build-planning#208

`xgboost` wheels currently follow this pattern:

* `xgboost` (CUDA 12 support by default)
* `xgboost-cu12` (always CUDA 12)
* `xgboost-cu13` (always CUDA 13)

See:

* dmlc/xgboost#10807 (comment)
* dmlc/xgboost#11677

This switches `cuml` to use those suffixed wheels, so the correct version will always be pulled in. As of this PR, `cuml-cu13[test]` will now depend on `xgboost-cu13` instead of `xgboost`.

## Notes for Reviewers

### XGBoost doesn't have arm64 wheels with CUDA 13 support

See #7550 (comment)

Authors:
  - James Lamb (https://github.com/jameslamb)

Approvers:
  - Bradley Dice (https://github.com/bdice)

URL: #7550

Author: jameslamb

dependencies.yaml

@@ -155,6 +155,7 @@ files:
       - py_run_cuml
       - py_version
       - test_notebooks
+      - test_python_xgboost
   py_build_cuml:
     output: pyproject
     pyproject_dir: python/cuml
@@ -506,10 +507,39 @@ dependencies:
           # included in the "devcontainers" key. The libxgboost package depends
           # on librmm but we do not want to have a package depending on librmm
           # in devcontainers since it should be built from source.
-          - &rapids-xgboost rapids-xgboost==26.2.*,>=0.0.0a0
+          - rapids-xgboost==26.2.*,>=0.0.0a0
+    specific:
       - output_types: [requirements, pyproject]
-        packages:
-          - &xgboost xgboost>=2.1.0
+        matrices:
+          - matrix:
+              cuda: "12.*"
+              cuda_suffixed: "true"
+            packages:
+              - xgboost-cu12>=2.1.0
+          - matrix:
+              arch: aarch64
+              cuda: "13.*"
+              cuda_suffixed: "true"
+            packages:
+              # TODO: always install xgboost-cu13 once there are xgboost-cu13 aarch64 wheels
+              #
+              # xgboost-cu13 originally shipped without aarch64 wheels
+              # Using the unsuffixed name pulls in CUDA 12 packages for some versions and may
+              # at some point in the future pull in CUDA 13 packages.
+              #
+              # ref: https://github.com/rapidsai/cuml/pull/7550#issuecomment-3603279934
+              - &xgboost_unsuffixed xgboost>=2.1.0
+          - matrix:
+              arch: x86_64
+              cuda: "13.*"
+              cuda_suffixed: "true"
+            packages:
+              - xgboost-cu13>=2.1.0
+          # this unsuffixed dependency is useful in environments where xgboost is built from source
+          # (like RAPIDS DLFW builds)
+          - matrix:
+            packages:
+              - *xgboost_unsuffixed
   test_notebooks:
     common:
       - output_types: [conda, requirements]
@@ -523,11 +553,9 @@ dependencies:
       - output_types: conda
         packages:
           - matplotlib-base
-          - *rapids-xgboost
       - output_types: requirements
         packages:
           - matplotlib
-          - *xgboost
   depends_on_cuda_python:
     specific:
       - output_types: [conda, requirements, pyproject]