Move element counts to parent nodes

[lorentey]

The tree currently stores the count of elements under each node directly in the nodes themselves. It would probably lead to faster offset-based navigation if these counts were moved to an array/buffer in the parent node instead.

We need to decide what to do with the root node's count.

An implementation of this would need to prove that it's an overall improvement, so it should come with benchmarking data — see issue #5.