Make B-Tree sorting blazingly fast

[komzpa]

• make gserialized_get_sort_hash function;
• make it return least possible value on EMPTY;
• make it read not box, but first coordinate of gserialized. This way no geography/geometry branching is needed, and it should be faster than getting a float-rounded box center for geometries that don't store it;
• make fast path for complete equality for fast GROUP BY in gserizlized_cmp;
• create sortsupport function that will employ gserialized_get_sort_hash and full equality fast path;
• investigate when to abort abbreviated compare in sort. (this is going to happen more often on 32bit machines, as sortsupport key will be 32bit, containing half of 64bit key).

[komzpa]

This ticket is spin off from #3864, it's not possible in 2.4 branch since it will change behavior and require opclass recreation.

[komzpa]

patch ​https://github.com/postgis/postgis/pull/436

[komzpa]

In 17600:

Fast Hilbert btree.

Requires btree indexes to be recreated.

Closes ​https://github.com/postgis/postgis/pull/436
Closes #3883