WIP: compute-matrix: add 'docs-build' matrix, switch docs builds to RTX PRO 6000

[jameslamb]

Introduces a docs-build matrix to the compute-matrix job, so we can centrally control which runner types are used for docs builds.

Details

RAPIDS projects generally need a GPU in their docs builds because things like Sphinx autosummary require actually import-ing the libraries, and many of the libraries require a GPU at runtime.

Today, projects are using the custom-job shared workflow and hard coding a GPU type, mostly L4.

  docs-build:
    uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@main
    if: fromJSON(needs.changed-files.outputs.changed_file_groups).build_docs
    with:
      build_type: pull-request
      node_type: "gpu-l4-latest-1"
      arch: "amd64"
      container_image: "rapidsai/ci-conda:26.10-latest"
      script: "ci/build_docs.sh"

(example from NVIDIA/cuml)

L4 runners are heavily used in the shared CI runner system we use right now, which can result in long queue lines:

Look how often they've been 100% utilized in the last 2 days:

We'd like to switch to a less-busy runner type, such as RTX Pro 6000:

Centralizing that choice of runner type would mean that changing it in the future wouldn't require PRs like NVIDIA/cudf#23878 in every repo. I think that's justified here... it's really unlikely that docs builds would be sensitive to the particular details of a runner, as long as it's a GPU RAPIDS supports.

[bdice]

Suggested change

	# Docs builds just need any GPU. Centrallizing the matrix reduces the effort to switch
	# Docs builds just need any GPU. Centralizing the matrix reduces the effort to switch

[bdice]

Capacity is more important here. Even if we queue sometimes, a runner type with high capacity will chew through its queue faster.

Suggested change

	# - some GPU with low queue
	# - some GPU with high capacity (low queue)

[bdice]

I'd like to use some version of CUDA 13 for our docs builds. Python version is less important. We have better coverage of the environments for CUDA 13 and fewer packaging oddities as CUDA 12 gets harder to support.