Previously, this was done by looking at the average of the predictions
over the posterior predictive. This was removed, as this didn't make
much sense.

The evaluate_tabular script was simplified to ease the analysis of more
quantities. Several coverage quantities and the bias were added to these
quantities. However, they perform rather poor.

Added EiV training scripts for the three datasets and moreover included
bias evaluation in `evaluate_tabular`. The inclusion of some coverage
measure is still needed.

For NLL metric, weird convention from MC Dropout was integrated.
Power Plant example missing.