For simulated examples the true coverage is evaluated. For all other
datasets create_tabular will show None.
The coverage was reversed to use the prediction (mean under the
posterior predictive). Note, that the "repeated sampling" assumption is
treated indirectly via averaging over the 10 seeds. For simulated data
this works decently, for real data one often (but not always, cf. wine)
sees a discrepancy.

The script evaluate_tabular.py was renamed to evaluate_metrics. This
script now not only prints the results, but also stores them in JSON
files in a Experiments/results folder (should be created). These files
can be read via a new script create_tabular.py
The JSON files have now all been changed to a less frequent update of
std_y.

Based now on JSON files in results folder. evaluate_tabular.py has been
renamed into evaluate_metrics. JSON files have also been updated. Need
to check whether correct now for all datasets.

Corrected eiv_prediction_number_of_draws in JSON configuration files
and updated the intervals of std_y updating for some of the datasets.

In the folder `Experiments/configurations` JSON files are included that
contain the configuration for training and evaluating. All training
scripts were replaced by two files `train_eiv.py` and `train_noneiv.py`
that read from this folder. The hitherto existent training
configurations from the training scripts were copied to the JSON files.
For `yacht_hydrodynamics` the configuration was updated.
The script `evaluate_tabular.py` now also reads from this folder.

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.