[FEA] Add post_block layer to MultiClassClassificationTask class

[rnyak]

🚀 Feature request

Currently we get logits as outputs if we train a MultiClassClassificationTask() model. It'd be useful to add an optional post-block layer in the MultiClassClassificationTask class and set it to softmax layer as default. That way users can generate predict probas as well.

[rnyak]

in the prediction mode we can make the softmax as default to get the probas.