Here is a second patch, quite independent of the first one, that gets rid of some other repetitive queries. On the regression database, the number of queries needed to do "pg_dump -s regression" drops from 3260 to 2589, and on my machine it takes 1.8 sec instead of 2.1 sec.
What's attacked here is a fairly silly decision in getPolicies() to query pg_policy once per table, when we could do so just once. It might have been okay if we skipped the per-table query for tables that lack policies, but it's not clear to me that we can know that without looking into pg_policy. In any case I doubt this is ever going to be less efficient than the original coding.