Switch to AMD HIP

[terhorst]

It shouldn't be too difficult to rewrite everything using ROCm/hip. The APIs are almost identical, and phlash does not rely on any CUDA-specific stuff. Then we could get acceleration on almost all existing GPUs (save Mac).