ciflow/torchtitan/186014: torchtitan: resolve build/test CUDA version from CUDA_STABLE

The torchtitan workflow hardcoded the CUDA version (13.0) in the build

environment string, the Docker image name, and the cuda-version input. Add a

get-cuda-version job that reads CUDA_STABLE via

.github/scripts/get_ci_variable.py --cuda-stable-version and exposes it as a

job output, then construct those values from that output so they track the

stable CUDA version instead of being pinned independently.

Note the Docker image must still exist in docker-builds.yml for the resolved

version, so a stable-CUDA bump remains gated on that image being built; this

change just removes the redundant hardcoded version from this workflow.

Test Plan:

Verified the helper resolves the expected version locally:

$ python3 .github/scripts/get_ci_variable.py --cuda-stable-version

13.0

With CUDA_STABLE=13.0 the build-environment resolves to

linux-jammy-cuda13.0-py3-gcc11 and the image to

ci-image:pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11, matching the previous

hardcoded values. Trigger ciflow/torchtitan to confirm the get-cuda-version

job output feeds the build/test jobs and they run on CUDA 13.0.

Authored by Claude.