Potential loose/packed ref race when deleting packed refs #5003

As @peff noted in #4981:

refdb_fs_backend__delete_tail doesn't have this bug, but I think creates its own race. It deletes the loose ref first, before touching packed-refs. During the moment between those two steps, a reader would see the packed-refs value as the authoritative one, but it's garbage at that point (it was some previous value that the ref had, but it could be very old, and possibly not even a reachable object). This was fixed in upstream Git by git/git@dc39e09.