More prominent warning for classifiers that do not support expected probabilities

For classifiers that do not support expected probabilities, we currently rely on SKLL to raise a warning and proceed generating integer scores. The final report still says "Predictions analyzed in this report are expected scores".
This is easy to miss when running multiple experiments.

We might consider either raising an error during model training or at least changing the wording in the report.

Start by tracing the model-training path and the final report wording, using SKLL's warning behavior as the current reference. Decide whether completion means failing during training or making the report clearly identify integer scores, then verify the chosen behavior across multiple experiments.

Scritto dal modello di indicizzazione a partire dal testo della issue.