Note we don't actually have to roll the dice for each byte, but continue doing the statistically equivalent but more performant thing we do now for allocations. This is more important now that bug 1068988 introduces sizes to the log, and will be doubly important after we add heterogeneous types to the allocations log.