Build extended stats on coordinators during ANALYZE

Build extended stats on coordinators during ANALYZE

When running ANALYZE on a coordinator, we simply fetch the statistics
built on datanodes, and keep stats from a random datanode (assuming all
datanodes are similar in terms of data volume and data distribution).

This was only done for regular per-attribute stats, though, not for the
extended statistics added in PostgreSQL 10, causing various failures in
stats_ext tests due to missing statistics. This commit fixes this gap
by using the same approach as for simple statistics - we collect stats
from datanodes and keep the first result we receive for each statistic.

While working on this I realized this approach has some inherent issues,
particularly on columns that are distribution keys. As we keep stats
from a random node, we completely ignore MCV and histograms from the
remaining nodes. That may cause planning issues, but addressing it is
out of scope for this commit.