Fix CI artifact download failures in sGPU/mGPU test jobs

[VeeraRajasekhar]

Fixes intermittent CI failures in sGPU Tests (mi30x) and mGPU Torch (mi35x) jobs caused by partial artifact downloads on slow self-hosted runners.

Root Cause

All four build outputs were combined into a single te-rocm-wheels artifact (~743 MB). GitHub Actions upload-artifact@v4 stores files alphabetically in a zip, so the archive order was:

1. transformer_engine-*.whl (~1 MB)
2. transformer_engine_rocm7-*.whl (~700 MB)
3. transformer_engine_rocm_jax-*.tar.gz (~20 MB)
4. transformer_engine_rocm_torch-*.tar.gz (~20 MB)

On slower self-hosted runners, downloading the 743 MB archive took 5–9+ minutes. The download-artifact@v4 action reported success even when the download was truncated, because the whl files (items 1 and 2) were extracted first. The tar.gz sdist files (items 3 and 4), which appear after the 700 MB whl in the zip, were never extracted. This caused the "Install packages" step to fail silently with an empty TE_FW_PKG variable.

Fix

Split the single artifact into two:

• te-rocm-wheels — .whl files only (~700 MB)
• te-rocm-sdists — .tar.gz sdist files only (~40 MB)

The sdists are now downloaded in a separate, independent step. Because the sdist artifact is small (~40 MB), it downloads quickly and completely on all runners regardless of network speed, preventing the partial-extraction failure.

Changes

• .github/workflows/rocm-wheels-build.yml: Replaced the single upload-artifact step with two steps — one for .whl packages (te-rocm-wheels) and one for .tar.gz sdists (te-rocm-sdists).
• .github/workflows/rocm-ci.yml: Added a Download sdist packages step in both sgpu_tests and mgpu_tests jobs to download the new te-rocm-sdists artifact alongside the existing wheel download.