This saves some memory from each index relcache entry. At least on a 64-bit
machine, it saves just enough to shrink a typical relcache entry's memory
usage from 2k to 1k. That's nice if you have a lot of backends and a lot of
indexes.