Re: Costing elided SubqueryScans more nearly correctly

I wrote:
> I also notice that setrefs.c can elide Append and MergeAppend nodes
> too in some cases, but AFAICS costing of those node types doesn't
> take that into account.

I took a closer look at this point and realized that in fact,
create_append_path and create_merge_append_path do attempt to account
for this.  But they get it wrong!  Somebody changed the rules in
setrefs.c to account for parallelism, and did not update the costing
side of things.

The attached v2 is the same as v1 plus adding a fix for this point.
No regression test results change from that AFAICT.