Multiclass Classification catalog does not contain EvaluateNonCalibrated

[artidoro]

The Multiclass Classification catalog has non calibrated trainers (SdcaNonCalibratedMulticlassTrainer) along with many calibrated trainers.

Currently there is only one Evaluate method, which assumes that the model is calibrated, and therefore produces LogLoss metrics. These metrics don't makes sense for the non calibrated trainers.

We should therefore add an evaluation method named EvaluateNonCalibrated for the non calibrated trainers.