Ensure compatibility with AMD, Metal, Intel (?) GPUs

[edoyango]

Currently the DynamicDet assumes only CPU or CUDA capabilities. i.e., when specifying --device, only cpu and device id are specified, where device id is assumed to be a CUDA GPU. Should be able to specify something like

AMD:

--device rocm:0

Metal:

--device mps # I don't believe there are setups with multiple Metal GPUs

But, if just a device index is specified, then it should also work e.g.

--device 0

Should choose the appropriate PyTorch device automatically.

[edoyango]

Looks like HIP uses the same semantics as CUDA https://pytorch.org/docs/stable/notes/hip.html. i.e., using --device 0,1,2,3 might work out-of-the-box for AMD. Still need to look at Metal + Intel.

AMD might not even affect AMP https://rocm.blogs.amd.com/artificial-intelligence/automatic-mixed-precision/README.html#torch-autocast

Metal (MPS) AMP looks to be WIP pytorch/pytorch#99272

Looks like Intel has AMP working: https://intel.github.io/intel-extension-for-pytorch/latest/tutorials/features/amp.html