review: table function support

Generally, the patch looks fine. There are few issues still:

- plpgsql: the result columns _do_ create local variables.
AIUI, they should not?

- pg_dump: is the psql_assert() introduction necessary, considering it
is used only in one place?

- There should be regression test for plpgsql too, that test if
the behaviour is correct.

- The documentation should mention behaviour difference from OUT
parameters.

Wishlist (probably out of scope for this patch):

- plpgsql: a way to create record variable for result row. Something like:

CREATE FUNCTION foo(..) RETURNS TABLE (..) AS $$
DECLARE
retval foo%ROWTYPE;

Currently the OUT parameters are quite painful to use due to bad
name resolving logic. Such feature would be perfect replacement.

--
marko