Fit photometry in flux space rather than magnitude space

[guillochon]

The primary benefit of fitting in flux space is that the flux errors are close to Gaussian and symmetrical as opposed to the asymmetrical, non-Gaussian errors in magnitude space. The following are some notes on the proposed changeover.

• Most photometry in the literature is presented in magnitudes, which means that fitting in flux space would involve converting the observed magnitudes to fluxes. In cases where the asymmetrical error bars are available in the data, there may be a way to recover the original fluxes and their errors.
• The GP and ML error models will simplify; we will no longer need to have two half Gaussians in the kernel for each point.
• This would allow us to more accurately do the SDSS asinh mags (see #24).
• Upper limits could be treated much more directly rather than the hacky way it is currently done (see #30).
• The conversion to magnitudes from the models is done almost at the end of the stack, meaning the model modules wouldn't need to change very much.
• Whether the light curve outputs are fluxes or magnitudes should be selectable via different output classes (i.e. FluxCurve and MagnitudeCurve classes rather than just LightCurve).