Transfer functions: specify number of gstar knots

[fjebaker]

It would be better to get the transfer functions to sample evenly in $g^\ast$ instead of in $\theta$.

The proposed algorithm change is

• estimate gmin gmax coarsely.
• for $g^\ast$ in set of target $g^\ast$, root find for radius, then root find over $\theta$ for the corresponding $g^\ast$.

An issue that might crop up here are those transfer functions where the inferred extrema are worse than the area extrema. In these cases we normally see an error of 1e-4 or so, but that could still be quite damaging to the calculation. Perhaps, however, we can spend more time in the extrema optimizer at the trade off of needing fewer $g^\ast$ than theta to sample the transfer functions fully.