And running CUDA binaries doesn't crash anymore, though they still do
not produce any meaningful coverage report, even for host code. This is
due to nvcc's temporary source files, we will have to compile with
clang++ natively to fix this.