Low density regions and splitting

[benjeffery]

Currently for large scale inference were splitting at regions of low, or no, site density. This results in some chrom arms with 3-4 tree sequences, which is annoying for downstream analysis. To over come this we either

1. Rejoin after inference
2. Tell tsinfer to terminate ancestor generation at the boundaries of drop out regions, this would mean that no edges span the bad regions.

It feels like option 2 is correct here and wouldn't be too tough to implement. Interested in others thoughts here. @jeromekelleher, @hyanwong