This still however removes `MemoryIterate' from allocations, because
it's a real performance pickle. From my tests parsing large sync
replies, this commit is near instant(pv reports 23MiB/s on a 2MB sync).
Still need a good compromise along MemoryIterate(like maybe find out a
clever way to only make it run on a small subset, or maybe just randomly
run it from time to time, or maybe just roll without it except on some
soft of debug mode????)
