[Build] Restore CUDA 13 Linux aarch64 core wheels

[bvolpato]

Describe the bug

transformer-engine-cu13 stopped publishing Linux aarch64 wheels after v2.16.0.

PyPI contains both platforms for v2.16.0:

• transformer_engine_cu13-2.16.0-py3-none-manylinux_2_28_aarch64.whl
• transformer_engine_cu13-2.16.0-py3-none-manylinux_2_28_x86_64.whl

For v2.16.1 and v2.17.0, PyPI contains only the x86_64 wheel. This makes installation on GB200/aarch64 fall back to a resource-intensive native build.

Steps to reproduce

Inspect published files for:

• https://pypi.org/project/transformer-engine-cu13/2.16.0/#files
• https://pypi.org/project/transformer-engine-cu13/2.16.1/#files
• https://pypi.org/project/transformer-engine-cu13/2.17.0/#files

Or attempt an ARM64 binary-only download:

python -m pip download \
  --only-binary=:all: \
  --platform manylinux_2_28_aarch64 \
  --python-version 3.12 \
  --implementation cp \
  --abi cp312 \
  --no-deps \
  transformer-engine-cu13==2.17.0

Expected behavior

Publish transformer_engine_cu13-<version>-py3-none-manylinux_2_28_aarch64.whl alongside the x86_64 core wheel for CUDA 13 releases.

ARM64 release automation also appears unhealthy:

• v2.16.1 ARM wheel job failed after exhausting runner disk space: https://github.com/NVIDIA/TransformerEngine/actions/runs/28210514784/job/83570631411
• v2.17 ARM wheel job failed before publishing an asset: https://github.com/NVIDIA/TransformerEngine/actions/runs/30397388710/job/90403758771

The v2.16.1 runner reported No space left on device while writing its diagnostic log. This may be separate from the core-wheel publishing path, but it shows the release currently lacks reliable ARM64 coverage.

Environment overview

• Platform: Linux aarch64
• Hardware: NVIDIA GB200
• Python: 3.12
• CUDA: 13.0
• Transformer Engine: 2.16.1 and 2.17.0