Simply symlinking doesn't actually quite work as evidenced by my CMake CI bot
(<https://open.cdash.org/test/498767666>), where I applied it manually.
Having a shim /usr/lib/cuda/include/cuda.h:
#pragma push_macro("CUDA_VERSION")
#define CUDA_VERSION 11030
#pragma pop_macro("CUDA_VERSION")
#include_next <cuda.h>
Solves the issue. This was suggested in the LLVM discussion
(<https://reviews.llvm.org/D108247#3025238>).
It is unclear to me why the symlink variant didn't work as previously when Clang
detected the version using version.h having the cuda.h simply in /usr/include
and thus the default search path was fine.
Though Clang is noted to be sensitive to the ordering of CUDA and other includes
due to it wrapping/shimming them to make them work with Clang.
