ROCm: Use correct device type when exporting tensors to DLPack (#40124)
Summary:
Before this PR, DLPack export was tricked by the CUDA masquerading of the HIP backend into thinking that it was exporting a CUDA tensor. We change that to use the ROCM device type instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40124
Differential Revision: D22145215
Pulled By: ezyang
fbshipit-source-id: 276f709861c55f499ae753d0bba48ddcc8b85926