Add Julia implementation of r2_score

[sethaxen]

This PR implements r2_score in Julia.

Computationally, the implementation is equivalent to Python's. However, it has a few extra convenience features:

• It takes output (draws[, chains], noutputs) for consistency with our other methods.
• It produces the output as a NamedTuple, similar to ess_rhat
• Similar to loo_pit, it has a method that takes an InferenceData, attempts to find the y and y_pred data in it, and then permutes dims as required by the default r2_score method.

[ahartikainen]

I know y and y_pred are magic, but I wonder if it should try to pair first variables (with or without _pred) in posterior and posterior_predictive groups, and then any xxx <--> xxx_pred

[sethaxen]

I know y and y_pred are magic, but I wonder if it should try to pair first variables (with or without _pred) in posterior and posterior_predictive groups, and then any xxx <--> xxx_pred

I implemented this in a new utility function observations_and_predictions, which is now used by both loo_pit and r2_score.