PM link

Now the cost of select({key}, offset) in TREE index with N tuples is O(log(N) + offset). That's a big problem for those users who use huge offsets. Actually it seems that it can simply optimized to O(log(N)) with a tiny cost - just store the total size of subtree in every node of bps tree.

On this way we could also optimize cost of count{key} to O(count).