Extend unit test for trainers

[accosmin]

• test_trainer_affine: check that all trainers can decrease the training error to epsilon3 with every optimization method
• test_trainer_classifier: same but for a simple classification task - classify an input as positive if its value is positive (+/- bias)

[accosmin]

May also write experimentation scripts to evaluate trainers for these synthetic datasets.

[accosmin]

In the end the affine task was extended to produce samples for either regression or multi-class classification.