According to buildfarm member sungazer, the behavior of VACUUM can be
unstable in these tests even if we prevent autovacuum from running on
the tables in question, apparently because even a manual vacuum can
behave differently depending on whether anything else is running that
holds back the global xmin. So use a temporary table instead, which
as of commit a7212be8 enables
vacuuming using a more aggressive cutoff.

This approach can't be used for the regression test that involves a
materialized view, but that test doesn't run vacuum, so it shouldn't
be prone to this particular failure mode.

Analysis by Tom Lane. Patch by Ashutosh Sharma and me.

Discussion: http://postgr.es/m/665524.1599948007@sss.pgh.pa.us