Improve hash_any() to use word-wide fetches when hashing suitably aligned

data.  This makes for a significant speedup at the cost that the results
now vary between little-endian and big-endian machines; which forces us
to add explicit ORDER BYs in a couple of regression tests to preserve
machine-independent comparison results.  Also, force initdb by bumping
catversion, since the contents of hash indexes will change (at least on
big-endian machines).

Kenneth Marshall and Tom Lane, based on work from Bob Jenkins.  This commit
does not adopt Bob's new faster mix() algorithm, however, since we still need
to convince ourselves that that doesn't degrade the quality of the hashing.

contrib/dblink/sql/dblink.sql

@@ -340,7 +340,15 @@ UNION
 (SELECT * from dblink_get_result('dtest3') as t3(f1 int, f2 text, f3 text[]))
 ORDER by f1;
 
-SELECT dblink_get_connections();
+-- dblink_get_connections returns an array with elements in a machine-dependent
+-- ordering, so we must resort to unnesting and sorting for a stable result
+create function unnest(anyarray) returns setof anyelement
+language sql strict immutable as $$
+select $1[i] from generate_series(array_lower($1,1), array_upper($1,1)) as i
+$$;
+
+SELECT * FROM unnest(dblink_get_connections()) ORDER BY 1;
+
 SELECT dblink_is_busy('dtest1');
 
 SELECT dblink_disconnect('dtest1');