Use `rocm-archs` in CMake

[danieldk]

We currently don't use them yet. Similar to CUDA, we should probably compute the intersection with all supported arch and then compile each kernel with the intersection.