Idea: no-store and no-load FFI calling convention annotations

[lukego]

Here is an idea to optimize calls to FFI code when we know that the call won't load and/or store C values that might be cached in JIT registers. This complements #231 to solve #230.

The idea is simply to declare when a load and/or store barrier is not needed as in:

void __noload foo();
void __nostore bar();
void __noload __nostore baz();

This would make calls to "pure" functions efficient e.g. to an assembler routine that loads the CPU timestamp from RDTSCP without touching memory.