atomic RMW intrinsics: avoid unnecessary ptr/int conversions

[RalfJung]

Currently, the type of out atomic RMW intrinsics looks like

fn atomic_xadd_seqcst<T: Copy>(_dst: *mut T, _src: T) -> T

However, this is not quite what we want: for atomic operations on a pointer, we want dst to be something like *mut *mut T, but src should be usize. The return type should be *mut T.

This would let us avoid some unnecessary casts in AtomicPtr, and shift the burden of mapping this operation to something LLVM supports into the backend. It also makes the semantics of these operations more clear: only the provenance of the in-memory data matters; src carries no provenance.