WIP patch for allowing COLLATE in plpgsql variable declarations

The attached code patch (seems to work, but no regression tests or doc
fixes yet) allows plpgsql variable declarations like

declare
x text collate "en_US";

Such a collation spec overrides the default variable collation that is
otherwise inherited from the function's input collation. I believe
this is essentially what we agreed to do a couple weeks ago.
Comments, objections?

regards, tom lane