add rocm pipeline

[AlbertThie]

Add a separate pipeline to use GPUs on a ROCM environment. This pipeline avoids using gpu-numba, as this has not been updated for ROCM for quite some time. The loading and unloading is a bit hacky, but I haven't noticed too much perfomance degradation.

Passed all pytest tests.

Tested on a node with 8 AMD Instinct MI250X GPU's.

Please note I did not have access to a NVIDIA multi GPU environment, some testing may be warranted.

[zubatyuk]

Thanks for the contribution.

Couple of issues:

I notice .cpu() and .cuda() transfers for the tensors.
Are these host-device transfers really required?
You should use current device index instead of the first one.

I still need to confirm that the PR does not brake CUDA and CPU compatibility.