Fix propagating attnotnull in multiple inheritance

In one of the many strange corner cases of multiple inheritance being
used, commit b0e96f3119 missed a CommandCounterIncrement() call after
updating the attnotnull flag during ALTER TABLE ADD COLUMN, which caused
a catalog tuple to be update attempted twice in the same command, giving
rise to a "tuple already updated by self" error.  Add the missing call
to solve that, and a test case that reproduces the scenario.

As a (perhaps surprising) secondary effect, this CCI addition triggers
another behavior change: when a primary key is added to a parent
partitioned table and the column in an existing partition does not have
a not-null constraint, we no longer error out.  This will probably be a
welcome change by some users, and I think it's unlikely that anybody
will miss the old behavior.

Reported-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: http://postgr.es/m/045dec3f-9b3d-aa44-0c99-85f6992306c7@gmail.com

src/backend/commands/tablecmds.c

@@ -7757,6 +7757,10 @@ set_attnotnull(List **wqueue, Relation rel, AttrNumber attnum, bool recurse,
 		List	   *children;
 		ListCell   *lc;
 
+		/* Make above update visible, for multiple inheritance cases */
+		if (retval)
+			CommandCounterIncrement();
+
 		children = find_inheritance_children(RelationGetRelid(rel), lockmode);
 		foreach(lc, children)
 		{