fitting the model.

In particular, the `learning_phase` flag will be automatically set to
False in case we are predicting on test data.

Moreover, a bug has been fixed causing labels preparation on
test data to be uneffective.

This should simplify the API when setting specific hyper-parameters using the dot notation

Now the DAP interface better differentiate in names
utility (private) methods - starting with `_`, from
public methods that has to be reimplemented by sub-classes.
This should make the API a little bit less awkward.
Moreover, now `deep_learning_dap` forces the settings of
`to_categorical` to True - which is True in the 99% of
the cases.

fixed bug with label_binarize adding trick to make it working also with Binary class probles, as expected