[FEA] Decouple this project from rapidsai/ci-imgs

[jameslamb]

Is your feature request related to a problem? Please describe.

Today, this project relies directly on the images built from https://github.com/rapidsai/ci-imgs

docker/Dockerfile

Line 18 in f49af7b

FROM rapidsai/ci-conda:${RAPIDS_VER}-latest AS dependencies

That has been convenient (for example, ci-imgs takes care of everything related to supporting new CUDA versions), but the tight coupling creates some development friction in both projects.

For example, see #831 for a description of how a "make images faster to pull in CI" change in ci-imgs resulted in a release blocker here in this repo.

Describe the solution you'd like

Nothing in this repo should depend on base images from rapidsai/ci-imgs.

Describe alternatives you've considered

N/A

Additional context

Changing base images is risky because it introduces possible changes in user-facing details like permissions, user IDs, etc.

I'd approach it like this:

1. inline everything from the Dockerfiles in into this repo with minimal changes https://github.com/rapidsai/ci-imgs
   • factor out rapidsai/ci-conda (remove reliance on rapidsai/ci-conda, use a versions.yaml #834)
   • factor out rapidsai/miniforge-cuda (remove reliance on rapidsai/miniforge-cuda (and therefore ci-imgs), upgrade cuvs-bench-cpu to miniforge 25.9.1-0 #836)
2. expand the container-canary checks in https://github.com/rapidsai/docker/tree/main/tests/container-canary to cover as much as we think is relevant
   • expand container-canary checks (Remove unnecessary Rocky Linux details, add docs, other small simplifications #839)
3. update Dockerfile here with optimizations specific to these images (for example, RAPIDS users probably do not need the gha-tools tools)
   • remove rockylinux-specific stuff (Remove unnecessary Rocky Linux details, add docs, other small simplifications #839)
   • move repetitive code into shared scripts (move more shared code into scripts #840)