In commit 9e3ad1aa I modified plpgsql
to use exec_stmt_return's simple-variables fast path in more cases.
However, I overlooked that there are really two different return
conventions in use here, depending on whether estate->retistuple is true,
and the existing fast-path code had only bothered to handle one of them.
So trying to return a scalar in a function returning composite, or vice
versa, could lead to unexpected error messages (typically "cache lookup
failed for type 0") or to a null-pointer-dereference crash.

In the DTYPE_VAR case, we can just throw error if retistuple is true,
corresponding to what happens in the general-expression code path that was
being used previously.  (Perhaps someday both of these code paths should
attempt a coercion, but today is not that day.)

In the REC and ROW cases, just hand the problem to exec_eval_datum()
when not retistuple.  Also clean up the ROW coding slightly so it looks
more like exec_eval_datum().

The previous commit also caused exec_stmt_return_next() to be used in
more cases, but that code seems to be OK as-is.

Per off-list report from Serge Rielau.  This bug is new in 9.5 so no need
to back-patch.