Fix run-tests CI import failures by aligning torch_scatter wheel with active torch ABI

[Copilot]

The run-tests GitHub Actions job failed during pytest collection due to a binary ABI mismatch between the container’s installed torch and torch_scatter. This change updates the test entry script to always install a torch_scatter wheel that matches the active torch version/backend before running tests.

• Root cause addressed

  • torch_scatter in the runtime environment was built against a different torch ABI, causing import-time symbol resolution errors.

• CI script hardening (ut.sh)

  • Detect active torch version at runtime.
  • Detect backend flavor (cpu or CUDA tag derived from torch.version.cuda).
  • Construct the matching PyG wheel index URL.
  • Force-reinstall torch_scatter from that index (--only-binary) before pytest.

• Resulting behavior

  • ut.sh no longer assumes the preinstalled torch_scatter is compatible with the current torch in the container.
  • Test collection uses a wheel aligned with the runtime torch ABI.

TORCH_VERSION=$(python -c "import torch; print(torch.__version__.split('+')[0])")
TORCH_BACKEND=$(python -c "import torch; print('cpu' if torch.version.cuda is None else 'cu' + torch.version.cuda.replace('.', ''))")
TORCH_SCATTER_INDEX="https://data.pyg.org/whl/torch-${TORCH_VERSION}+${TORCH_BACKEND}.html"
pip install --no-cache-dir --force-reinstall --only-binary=torch_scatter torch_scatter -f "${TORCH_SCATTER_INDEX}"