[CI] Use stable torch for the ROCm leg (#2910)

ROCm was the only leg in the test matrix on torch nightly; CUDA-auto and Metal
both use stable. That asymmetry meant an upstream PyTorch nightly regression
turned the AMD leg red on unrelated PRs, reading as self-hosted runner flakiness
rather than upstream churn.

Verified on the gfx942 runner against eb2136da. Stable resolves to
torch 2.13.0+rocm7.2 and gives results identical to nightly:

    1910 passed, 1237 skipped, 0 failed

Side benefit: each new nightly pulled ~16G of wheels into the runner's uv cache,
which reached 48G in two days. Stable removes that churn.