Add aarch64 variants for ghci-snl gnu-cpu-env and gnu-cuda-env

[Copilot]

ghci-snl only built x86_64 images. This adds native aarch64 equivalents: gnu-cpu-env for macOS M1/Apple Silicon, and gnu-cuda-env for NVIDIA aarch64 "superchip" systems (Grace Hopper/Blackwell), built on GitHub's native ubuntu-22.04-arm runners (no QEMU).

Base image

• SPACK_ARCH is now a build ARG in base/Dockerfile instead of hardcoded, defaulting to linux-rhel9-x86_64 (unchanged behavior). aarch64 builds override it:

  docker build --build-arg SPACK_ARCH=linux-rhel9-aarch64 ...
  

gnu-cpu-env-aarch64

• No Dockerfile changes needed — compiler/Dockerfile and final/gnu-cpu-env/Dockerfile already parameterize on ${SPACK_ARCH} generically. The workflow now builds this same Dockerfile a second time for aarch64.

gnu-cuda-env-aarch64 (new)

• New final/gnu-cuda-env-aarch64/Dockerfile. spack@0.23.1 (pinned repo-wide) only packages CUDA up to 12.6, so CUDA is installed from NVIDIA's official RHEL9/SBSA repo instead of spack install cuda@..., landing on CUDA 13.0.
• Uses the prebuilt cupy-cuda13x wheel and torch==2.13.0+cu130 (vs. x86_64's source-built cupy against spack's CUDA).

CI workflow

• .github/workflows/ghci-snl.yaml: base, compiler, and final-env jobs are now matrixed over arch, adding base-aarch64, gnu-13.3.0-aarch64, gnu-cpu-env-aarch64, and gnu-cuda-env-aarch64 tags. Intel oneAPI has no aarch64 support, so intel-cpu-env stays x86_64-only.

Docs

• ghci-snl/README documents the new tags and the aarch64 base-image build command.

Possible follow-ups (not included here)

• gnu-cuda-env/Dockerfile declares an unused ARG NVIDIA_DRIVER_VERSION=535; left as-is, out of scope.
• A future Spack 1.x migration would let CUDA 13 install via spack install cuda@13 directly, unifying the install path across arches — deferred since Spack 1.x deprecates the packages:all:compiler: model used repo-wide and needs broader testing.