[libc] Partially implement `atexit` on the GPU

The `atexit` function controls registering functions to call at the end
of the program. This is difficult to do in general on the GPU because of
the lack of a real mutex implementation. We primarily provide this for
testing where we can explicitly restrict how the `atexit` registration
functions are called. So we simply create a passthrough Mutex to get
past the usage of it as per @sivachandra's suggestion.

Depends on D149225

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D149226