CMAKE_CUDA_HOST_COMPILER cannot be set from a toolchain file
When cross-compiling a CUDA project with CMake's native CUDA support (version 3.11.0), I need to specify the CMAKE_CUDA_HOST_COMPILER so that nvcc can use the cross-compiler as the host compiler. The correct location to specify this would probably be the toolchain file, where the cross-compiler gets selected in the first place. This does not work, though, as the variable is always empty in the resulting build directory. It only works, when CMAKE_CUDA_HOST_COMPILER gets specified directly from the command line.