Add RMSEWithUncertainty likelihood to catboost model

[jonasracine]

Currently CatBoost fails with RMSEWithUncertainty as likelihood when output_chunk_length is greater than 1.

Crux of the problem: because CatBoost outputs two values [mu, sigma], arrays need to be reshaped all over the place. Contrary to what we thought, it seems that MultiOutputRegressor does not fail.

Some (messy) work has already been done and can be found on the branch feat/rmse_uncertainty_catboost

A good place to start debugging is to look at the following tests:

• test_models_accuracy_univariate
• test_models_accuracy_multivariate
• test_models_accuracy_multiseries_multivariate