One thing I forgot to mention about this experiment: when used in a
backend I think this probably needs a chunk size (what size?) and a
CFI(). Which is a bit annoying, because for best results we want
SSIZE_MAX, but in the case that copy_file_range() falls back to raw
copy, that'd do I/O work bounded only by segment size :-/ (That's not
a problem that comes up in the similar patch for pg_upgrade.)