I neglected to test this scenario while preparing commit f3faf35f,
so of course it was broken, thanks to some very obscure and undocumented
code in pg_dump.  Pre-v12 databases might have toast tables attached to
partitioned tables, which we need to ignore since newer servers never
create such useless toast tables.  There was a filter for this case in
binary_upgrade_set_type_oids_by_rel_oid(), which appeared to just
prevent the pg_type OID from being copied.  But actually it managed to
prevent the toast table from being created at all --- or it did before
I took out that logic.  But that was a fundamentally bizarre place to be
making the test in the first place.  The place where the filter should
have been, one would think, is binary_upgrade_set_pg_class_oids(), so
add it there.

While at it, reorganize binary_upgrade_set_pg_class_oids() so that it
doesn't make a completely useless query when it knows it's being
invoked for an index.  And correct a comment that mis-described the
scenario where we need to force creation of a TOAST table.

Per buildfarm.