This PR adds a new transform op that replaces `memref.alloca`s with `memref.get_global`s to newly inserted `memref.global`s. This is useful, for example, for allocations that should reside in the shared memory of a GPU, which have to be declared as globals.