In reply to @treeform "hippo also does nim": In this case,
- I wanted to avoid compiling through C++ and nvcc because for a library it
breaks custom flags, `-flag` has to become `-Xcompiler flag`.
- You force Nvidia/AMD compiler for the whole app and all its dependencies
instead of just the GPU modules, and dealing with the path differences was also
annoying in the past.
- The should ease porting to different backends like WebGPU, Vulkan, OpenCL,
Metal that have slightly different programming models than Cuda/HIP