Switch to radix sort?

[dgryski]

The sorting in the finalize step is very expensive. We can probably do better than quicksort since we know we're sorting integers. Especially if we want to update the tables and call Finalize() again, it would be nice not to hit O(n^2) behaviour (although Go drops to mergesort if the recursion level gets too high, we're still in O(n log n) territory).

[dgryski]

Double memory on radix sort is expensive though, especially at the scale we're looking at.

[dgryski]

If we halve the memory requires via #10, this is painful, but less so.

[dgryski]

Sorting the tables in serially reduces memory requirements, especially if we can reuse the temporary array.