implement second bitmap, ~2x speedup for train

[a10y]

^ this is a sad flamegraph. bzero is not a fun place to be spending 60% of your time.

How did we get from one to two?

1. Avoid initializing our counts1 and counts1 vectors
2. Implement a second bitmap index that limits our outer-loop iterations in optimize. Because counts1 and counts2 are not initialized, we check the bitmap before all accesses

This also gives us another 2x speedup on the train benchmark, which is nice