AMDGPU: Add patterns for i32/i64 local atomic load/store

Not sure why the 32/64 split is needed in the atomic_load
store hierarchies. The regular PatFrags do this, but we don't
do it for the existing handling for global.

llvm-svn: 335325