[mlir][memref] Implement lowering of memref.copy to llvm

This lowering uses a library call to implement copying in the general case, i.e.,
supporting arbitrary rank and strided layouts.