[Feature] Add benchmark for AMD accelerators

[NellyWhads]

AMD accelerators do not have competitive training and inference performance with NVIDIA accelerators. This lack of performance, in part, could be significantly improved with kernel level optimizations for pytorch.

This benchmark could serve as a community driven development accelerator to bridge this performance gap.

[simonguozirui]

Hi Nelly, thanks for taking an interest in the KernelBench project and the suggestion.

We recently spoke with the AMD team. A good first goal is to add HiP (AMD equivalent of CUDA) support as a KernelBench backend.

See efforts such as #37 to integrate AMD accelerators as a backend for KernelBench. We welcome contributors from AMD and beyond (from the community).