We're considering changing the implementation of targetlist SRFs
considerably, and a lot of the current behaviour isn't tested in our
regression tests. Thus it seems useful to increase coverage to avoid
accidental behaviour changes.

It's quite possible that some of the plans here will require adjustments
to avoid falling afoul of ordering differences (e.g. hashed group
bys). The buildfarm will tell us.

Reviewed-By: Tom Lane
Discussion: <20160827214829.zo2dfb5jaikii5nw@alap3.anarazel.de>