Separate the modelling from the likelihood functionality

[tilmantroester]

The way firecrown is structured currently strongly assumes that one wants to calculate a log likelihood. Often we only want to make some model predictions, which do not depend on the likelihood, data vectors, or covariances.

compute_theory_vector should only depend on things actually required to evaluate the theory vector. So in the case of two-point statistics, the n(z) and ell/theta ranges/binning, but not for example the covariance.

A step further would be to provide some utility to bootstrap a minimal sacc object from some more basic quantities, like type of statistic, a list of n(z), and ell bins, so that the user doesn't have to deal with sacc files at all if they just want a model prediction.

This is related to the discussion in #308 on standardising model specifications.