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.