The time varying transform models need refactoring

[Cadair]

We now have 12 model classes for handling all the permutations of varying celestial transforms (slit/not slit, inverse and 1-3D). This is insane.

Since these classes were originally written we have a fully working ND Ravel model which we should be able to use to flatten the lookup table axis to always 1D. This should mean at minimum we can get rid of the 2D and 3D classes.

This drawing illustrates what I envisage a compound model for a 2D slit variant would look like.

The main complication of this approach is that the inverse transform needs to be modified from the direct inverse of the forward transform (as illustrated by dashed lines).

I think the best approach for this would be to have a function which builds the forward and inverse transforms and combines them into a single CompoundModel class to return.

I think this should get us to the situation where we have at least no 2 and 3D variants of the classes and maybe the slit variants can be combined with some trickery in the model construction function.