Fix pg_dump to ensure that a comment on a table CHECK constraint cannot

Fix pg_dump to ensure that a comment on a table CHECK constraint cannot
be emitted too soon.  The previous code got this right in the case where
the CHECK was emitted as a separate ALTER TABLE command, but not in the
case where the CHECK is emitted right in CREATE TABLE.  Per report from
Slawomir Sudnik.

Note: this code is pretty ugly; it'd perhaps be better to treat comments
as independently sortable dump objects.  That'd be much too invasive a
change for RC time though.