Support CUDA mixed host/device source files

[jeroenk]

CUDA mixed host/device source files are much better supported in Clang nowadays, it might be worth revisiting whether this can be supported.

[jeroenk]

A possible approach is to pull in the libcxx C++ library and the musl C library. Then, pass -nostdinc to clang and also pass the locations of the libcxx and musl include files:

• For libcxx this means the include directory that occurs in the root of the repository.
• For musl this also means the include directory that occurs in the root of the repository. In addition, for 32-bit, and 64-bit pointers, pointers also arch/i386, and arch/x86_64, should be passed, respectively.

For the libraries to work correctly, a variation on the following will be required.

• libcxx: define __ELF__ and _GNU_SOURCE. Add definitions of _LIBCPP_HAS_THREAD_API_PTHREAD and _LIBCPP_HAS_MUSL_LIBC to the start of __config.
• musl: generate a bits/alltypes.h based on arch/.../bits/alltypes.h.in. This depends on the pointer width.

Besides the above, we will also need prototypes for CUDA runtime API functions (see runtime_patch.txt for a very minimal version).