Cross Validation Refactor

[akleeman]

Moves cross_validated_predictions to a method in RegressionModel which allows specialization of the cross validation. The only real change in behavior is the ability to obtain a cross validated MarginalDistribution. The existing code only allows for the cross validated mean.

The primary reason for this reorganization is to be able to have per model specialization of some of the cross validation routines. For example gp_fast_loo_nll which takes advantage of a linear algebra trick to accelerate leave one out predictions for Gaussian processes is now directly available using the RegressionModel api (instead of the current static cast hack.)

The first commit is just a lot of moving things around, the second contains the RegressionModel api changes. A followup PR will switch gp_fast_loo_nll to be a specialized method in gp.h.

[kmdade]

the trailing underscore convention is backwards here from what we normally have. I'm going to assume its consistent with the rest of albatross however which is all that really matters. I think we talked about this ages ago... :)

[akleeman]

yeah, the rest of albatross uses a trailing underscore to avoid shadowing. I started to try to change them all ... then decided it was more work than it was worth.

[kmdade]

Does this have to be templated? can it not just be hardcoded to use size_t or unsigned or whatever?

[akleeman]

Yeah this might be overuse of templating, in general the value type of col_indices can be anything that can be converted to an Eigen::Index, but now I think the only thing that is really used is s32. Whatever the decision is the same argument would apply to all the rest of the subset functions in this file as well.

[kmdade]

quick glance over, looks good to me. I will come back and check it out more thoroughly when i have a sec.

[seth-swiftnav]

lgtm