Previous coding was passing the wrong table's tuple descriptor, which
accidentally fails to fail because no existing test case exercises a
foreign key in which the referenced attributes are further to the right
of the referencing attributes.

Add a test so that further breakage is visible.

This got broken in 16828d5c.

Discussion: https://postgr.es/m/20180403204723.fqte755nukgm42uf@alvherre.pgsql