Split up Runtime Files and expose simpler FFI headers

[enricozb]

The C and CUDA runtime files are monoliths right now which make them a bit difficult to navigate. Additionally the FFI header file must be manually kept in sync. We should have smaller files, and and aim for an FFI header file that simply imports the headers we want to expose. Something like
ffi.c:

#include <runtime.h>
#include <io/result.h>
#include <io/readback.h>