fix slow filters on multi-valued properties in joined tables

[cportele]

Filters on values in a joined table (e.g. a value array with A_CONTAINS or A_OVERLAPS, or a negation of such a filter) took seconds to minutes on large tables when the value is frequent, because the planner had to materialize every matching row of the whole table:

• A_CONTAINS with a single value and A_OVERLAPS no longer add a GROUP BY without HAVING to the IN semi-join; it does not change the result, but it kept PostgreSQL from turning the subquery into a semi-join that probes the joined table per candidate
• A_CONTAINS with several values is the conjunction of one semi-join per value instead of a single subquery with GROUP BY ... HAVING count(distinct ...) = n; this also matches a list with duplicate values
• the negation of a single semi-join is NOT EXISTS over the unchanged subquery instead of NOT (... IN (...)), which the planner can only evaluate per candidate or hash; it is the same predicate, as the key of the main table is never null, and the planner turns it into an anti-join

The derived table of NOT EXISTS has an alias without AS, which Oracle does not accept there.