The aggregate-order difference explained in my previous commit
turns out to also affect the order of log entries emitted in the
contrib/sepgsql regression test.  Per buildfarm.

Discussion: https://postgr.es/m/21272.1563318411@sss.pgh.pa.us