Improve ancestor building by using more information

[hyanwong]

It looks like #903 is being caused by bad choices of ancestral vs derived states in the ancestor builder. We do have more information for building, e.g. we know after the first inference (or if we use the PBWT) whether we think that many of the haplotypes contributing to the consensus form a cluster of recent close relatives.

We should think about how to use this information to improve ancestor reconstruction. @petrelharp suggested building a "decision tree" rightwards and leftwards from each site, branching whenever we make a choice of 0 or 1 for an adjacent site, and picking e.g. the longest path. He pointed out that this can be (mostly) kept between ancestors, so should be efficient.