consider using a two-parameter lcfit? about sts

[cmccoy]

See here for point choosing / fitting: https://github.com/matsen/pplacer/blob/294-likelihood-fit/lcfit_src/lcfit.ml#L138-L281
I was just evaluating the likelihood on an evenly spaced grid and dropping low-likelihood areas.

Points chosen to to evaluate impacts lcfit on a single branch quite a lot.
Choosing a grid of points for the three parameter version was also seemed difficult.

My simulations are on stoat in /loc/ssd/cmccoy/lcfit-validation

We got decent estimates of marginal probability, but (relatively) small errors led to wildly different posterior probabilities:

Rejection sampling in sts might be better.