plpgsql: remove useless distinctions between record and row cases

I've been fooling around with rewriting plpgsql's composite-variable
handling, along the way to getting it to handle domains over composite.
I noticed that there's some really unnecessary complication in places
where it insists on separating "row" variables from "record" variables.
More usually, we'd handle that by using a generically-typed pointer and
then doing a node-type check where it's actually necessary to distinguish;
which it turns out is exactly one place, exec_move_row(). So attached is
a simple patch that eliminates the duplicative coding.

Barring objection, I'd like to push this shortly.

regards, tom lane