Now subdomains keep pending points in their interior, and we only
*split* subdomains when we get new data.

We need more than just a RB tree/skip list because we need to update
the priority of elements *without knowing their priority*. This
implies that we also need a hashmap.

The parallel algorithm is going to become more complicated now that
we can add points in subdomains without splitting.

Rearrange the paragraphs so they make logical sense.
Only include pseudocode for the parallel version of the algorithm.

This reverts commit 93e1de31.