Scale exponential prior on roots using root spans

[hyanwong]

Currently we are a bit biased in the most distant past, which we think is probably down to the exponential prior on the unconstrained (root) nodes. We could take advantage of the fact that we have a span on the roots to constrain the prior somehow. @nspope says:

span and age are roughly log-linear under the coalescent, but of course the exact relationship depends on demographic history, sequence boundaries, etc and is also likely distorted by tsinfer's heuristics. So, I think the trick is how exactly to incorporate span. I was thinking this could be estimated via the same EM algo that is currently used (that is, we specify a functional form mapping span onto scale, and use EM to fit it after each round of message passing).