New observation structure

[pmelchior]

Christmas PR!

This implements the suggestions in #220. In particular, it defines Observation as a Frame with a data payload and a log_likelihood method. The actual rendering method is provided by subclasses of Renderer. This avoids having to change the type of observation. Instead, Observation.match sets/changes the renderer.

As an additional bonus, this branch derives Renderer from Model, reflecting the fact that the mapping from model space to data space can include optimizable parameters (WCS, PSF, mag zeropoint #212, wavelength calibrations...). So, Renderer creates the parameterized transformation function and then applies it to the given model. This will serve us well for cases where we'd like to fit for the PSF in a crowded field etc.

Finally, this branch builds off prerender which is currently under review in #227. I suggest to do that PR first, and then move to this one, otherwise there are too many moving parts.

[herjy]

Did you branch this off of deconv_init and are trying to merge to master? I will review this after we merge 227 as the diff is currently huge and largely dominated by 227.

[pmelchior]

That's what I suggested

[pmelchior]

A few comments:

• ResolutionRenderer is a straight copy from LowResolutionObservation
• For Observation.data I was thinking of spectroscopic data sets, but this is a little premature. But I'd rather make this change now.
• The prerender option would need to move to the Renderer (which can decide if it is even possible to create), but I have removed the deconvolution init because it seems to not help. Since the code is in master, we can bring it back if we want to.