Test ability to detect wrongly polarised ancestral states in tsinfer

[hyanwong]

We might expect wrongly polarised ancestral states to produce an identifiable signal the dates generated by tsdate. We could test this both on simulated data (deliberately mispolarise some ancestral states, probably weighted towards higher frequency variants), or on real data (do an inference of real data using REF as the ancestral state, and see if those sites known, via other methods, to have an ALT ancestry, have an unusual dating pattern). I guess we might expect these to be cases where there are large conflicts between topology and mutational information (i.e. big areas with low mutational probability).