Commit 7cb0a423 overlooked that the multi-level partition test table
pagg_tab_ml still had an exactly even row split at its upper level of
partitioning, so that some of the sub-aggregation plan steps still had
exactly equal costs, leading to plan instability.  Tweak the partition
boundaries some more to make the row distribution unequal at both
levels.  This leads to more changes in the "expected" plan order than
the previous round, but it seems fine.  (Actually, I'm surprised that
this didn't affect even more plans in this test: looking at the
underlying costs shows that some of the parallel plan groups are
*not* getting sorted by cost.  Bug?)

Per buildfarm member lousyjack,
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lousyjack&dt=2020-04-04%2021%3A03%3A04

Discussion: https://postgr.es/m/24467.1585838693@sss.pgh.pa.us