Follow-up from #12908

This PR extends the join filters that get generated by hash joins to include an IN filter when the hash table is small enough. Rather than generating only a min-max filter, we generate an IN filter with all values in the hash table. This can greatly improve performance over the min-max filter when there are few values that are far apart. For example, if the hash table contains the values 1 and 100000, the IN filter might be much more effective as we can prune many more row groups.

The threshold for which we generate an IN filter is determined by the dynamic_or_filter_threshold setting, and defaults to 50 rows. The IN filter is pushed as an OPTIONAL_FILTER, which is currently only evaluated for zone-map pruning. As a result, the performance impact of pushing this filter is minimal, while the performance improvement from extra zone-map pruning can be significant.

Benchmark

Below is a benchmark that we run over TPC-H SF10 in which we run a join that is generated through an IN clause - we join on the min and max values of l_orderkey (meaning the generated min/max filters will not be effective in pruning rows).

SELECT *
FROM lineitem
WHERE l_orderkey IN (
    SELECT UNNEST([MIN(l_orderkey), MAX(l_orderkey)])
    FROM lineitem)
ORDER BY ALL;