Make fitting functions more generic

[cgarling]

I started with a few basic AMRs but as I start to consider more models it would be useful to formalize a generic interface. The three moving components in the hierarchical models are

1. the relation between the SFRs and the stellar metallicities (consider the difference between age-metallicity relations and mass-metallicity relations),
2. the parametric form of the stellar metallicity evolution, and
3. the model for metallicity dispersion at fixed time.

Part 1 sets the general form for the partial derivatives of the objective with respect to the SFRs -- i.e., for all AMR-type models we could probably use one fit_templates method, if we extract the other parts out.

Part 2 could be abstracted as something like a MetallicityModel type. Need to be able to compute mean metallicity as a function of time, or stellar mass, or whatever other dependent variable you have. Also need gradient -- partial derivatives with respect to fundamental variable and all fittable parameters.

Part 3 could be abstracted as something like a MetallicityDispersion type. Given a mean metallicity and a set of template metallicities, need to compute the relative weights of each. Also need the gradient of the relative weights -- partial derivatives with respect to fundamental variable and all fittable parameters.

Not sure this is worth the time for AMRs but it's more interesting in the case of mass-metallicity relations because there are more parametric forms we may be interested to try, and such a generic setup would require some upfront work to write but would make adding more models later considerably easier.

[cgarling]

Flatten.jl might be of interest for flattening model structs into formats that can be fed to optimizers.

[cgarling]

Fixed by #58.