Migrate to NVKS for amd64 CI runners

[bdice]

This migrates amd64 CI jobs (PRs and nightlies) to use L4 GPUs from the NVKS cluster.

xref: https://github.com/rapidsai/build-infra/issues/184

[jakirkham]

Is there something still needed for CUDA 12.8 here?

Seeing the following error on CI:

Traceback (most recent call last):
  File "/opt/conda/bin/rapids-dependency-file-generator", line 10, in <module>
    sys.exit(main())
             ^^^^^^
  File "/opt/conda/lib/python3.12/site-packages/rapids_dependency_file_generator/_cli.py", line 125, in main
    make_dependency_files(
  File "/opt/conda/lib/python3.12/site-packages/rapids_dependency_file_generator/_rapids_dependency_file_generator.py", line 474, in make_dependency_files
    raise ValueError(f"No matching matrix found in '{include}' for: {matrix_combo}")
ValueError: No matching matrix found in 'cuda_version' for: {'cuda': '12.8', 'arch': 'x86_64'}

Maybe we need to resolve the forward merger: #6272

[dantegd]

@bdice the current failures are because the jobs are picking a build of cudf nightly that doesn't have the fixes that 25.02a358 has

[jameslamb]

The most recent cuDF branch build finished a couple minutes ago (https://github.com/rapidsai/cudf/actions/runs/13078947512). I've merged brnach-25.04 in here to restart CI and pull in those new packages.