Proposed patch for ANALYZE overcounting dead rows

There's been a thread in pgsql-performance about ANALYZE reporting an
overly large number of dead rows because it counts anything that's not
visible-per-SnapshotNow-rules as "dead", in particular
insert-in-progress tuples are reported as "dead". This seems a bad
idea because it may prompt autovacuum to launch a useless vacuum.

The attached patch revises ANALYZE to distinguish tuple states the same
way VACUUM does, and to not count INSERT_IN_PROGRESS nor
DELETE_IN_PROGRESS tuples as either "live" or "dead".

Comments? In particular, should we report a count of in-doubt tuples in
the eventual LOG message, and does anyone want to argue for still
counting DELETE_IN_PROGRESS as "dead"?

regards, tom lane