Test and robustify shapelet PSF approximations

The CModel code ported from HSC only works as well as the ShapeletPsfApproximation algorithm that runs before it, but we've switched on the LSST side to a more flexible algorithm that isn't as nearly as battle-tested as what's been running on the HSC side, and there are some concerning indications from Perry Gee's work that it can be catastrophically slow on some reasonable PSFs. On this issue, I'll run it on some real HSC data and try to improve it, even if that means reducing the flexibility back to what was on the HSC side in some ways.