Fix indexable-row-comparison logic to account for covering indexes.

indxpath.c needs a good deal more attention for covering indexes than
it's gotten.  But so far as I can tell, the only really awful breakage
is in expand_indexqual_rowcompare (nee adjust_rowcompare_for_index),
which was only half fixed in c266ed31.  The other problems aren't
bad enough to take the risk of a just-before-wrap fix.

The problem here is that if the leading column of a row comparison
matches an index (allowing this code to be reached), and some later
column doesn't match the index, it'll nonetheless believe that that
column matches the first included index column.  Typically that'll
lead to an error like "operator M is not a member of opfamily N" as
a result of fetching a garbage opfamily OID.  But with enough bad
luck, maybe a broken plan would be generated.

Discussion: https://postgr.es/m/25526.1549847928@sss.pgh.pa.us