The work for this was done on DM-14574, but I've rebased it over to here to make the review of this and that easier.

Jim Bosch: do you think you would be able to review this? I extracted the code that incorporated the model uncertainty from DM-14574, to make the reviews less onerous; some of the preparatory code for that is still in this ticket, because I couldn't separate the cleanups from the actual fixes.

The bug in the photometry uncertainty calculation that I thought I'd found turns out not to exist (both the simple and constrained models end up with the flux and fluxErr calculations taking exactly the same form). Instead, I added the model uncertainty, which we should be including as part of the chi2. It doesn't shift the test calculations much, and it doesn't seem to help the HSC PDR1 data, but we might as well get it right.

Jenkins run: https://ci.lsst.codes/blue/organizations/jenkins/stack-os-matrix/detail/stack-os-matrix/28120/pipeline

What problem is this trying to solve now?  The only uncertainty in the chi^2 should be the measurement errors in the MeasuredStars; models don't have uncertainties.  The best-fit model parameters will have uncertainties, of course, and those need to be propagated when calibrating fluxes, but that should not affect the fit itself.

Marking Invalid.

After discussion with Jim Bosch, we decided that this change is not correct: the model uncertainty comes from the covariance of the measured uncertainties, so it is already incorporated in the fitter.