Perform multiple searches in parallel

With a binary tree we can a.) switch to a branchless form and b.) do the searches with a baked in prefetch call. We can then batch 4-8 array searches in parallel, and take advantage of the prefetches from #5.

With Eytzinger order #4, however, such extra prefetches would be random main memory reads if we needed to do extra levels. This would bound our parallelism by the number of structures at the same level as us. (2-3x currently.)

If the payout is large enough, the latter may supply motivation for considering a larger radix number system.