Use rapids-pip-retry in CI jobs that might need retries

[gforsyth]

Uses a retry wrapper for pip commands to try to alleviate CI failures due to hash mismatches that result from network hiccups

xref rapidsai/build-planning#148

This will retry failures that show up in CI like:

   Collecting nvidia-cublas-cu12 (from libraft-cu12==25.2.*,>=0.0.0a0)
    Downloading https://pypi.nvidia.com/nvidia-cublas-cu12/nvidia_cublas_cu12-12.8.3.14-py3-none-manylinux_2_27_aarch64.whl (604.9 MB)
       ━━━━━━━━━━━━━━━━━━━━━                 350.2/604.9 MB 229.2 MB/s eta 0:00:02
  ERROR: THESE PACKAGES DO NOT MATCH THE HASHES FROM THE REQUIREMENTS FILE. If you have updated the package versions, please update the hashes. Otherwise, examine the package contents carefully; someone may have tampered with them.
      nvidia-cublas-cu12 from https://pypi.nvidia.com/nvidia-cublas-cu12/nvidia_cublas_cu12-12.8.3.14-py3-none-manylinux_2_27_aarch64.whl#sha256=93a4e0e386cc7f6e56c822531396de8170ed17068a1e18f987574895044cd8c3 (from libraft-cu12==25.2.*,>=0.0.0a0):
          Expected sha256 93a4e0e386cc7f6e56c822531396de8170ed17068a1e18f987574895044cd8c3
               Got        849c88d155cb4b4a3fdfebff9270fb367c58370b4243a2bdbcb1b9e7e940b7be

This PR also updates the build images used for ci-wheel to use CUDA 12.8.0 (we no longer maintain CUDA 12.0.1 build images) and uses NVKS L4 GPUs for amd64 testing, to help move us off of the old CI cluster.

[bdice]

This is showing the same issues with running Numba tests on ARM that we've seen elsewhere. Let's wait to merge until that issue is resolved. cc: @gmarkall

[bdice]

This fails because we only build ci-wheel for the latest CUDA version of each major version. This repo uses CUDA 12.0.1 which is not a recent build of the ci-wheel image. I think the best fix here is to update the image tags to 12.8.0. We don’t plan to maintain CUDA 12.0.1 CI images.