[DO NOT MERGE] [ROCm] device sync on pytorch module exit

[dnikolaev-amd]

Failed test test_cuda.py::TestBlockStateAbsorption::test_no_triton_on_import

Failed in release/2.12 and upstream/main with "Segmentation fault (core dumped)"

Reproducer:

python -c "import torch; t=torch.eye(2, device='cuda'); print('ok')"

ok
Segmentation fault (core dumped)

This PR was made to save the solution found. It may not be needed.

[rocm-repo-management-api]

Jenkins build for 017382841d6df6a8704bce2240fe25f193dee5e8 commit finished as FAILURE
Links: Pipeline Overview / Build artifacts / Test Results