Generalize multiclass to support multilabel classification

[jasonnance]

Feature

Support multilabel classification instead of just multiclass classification.

Motivation

This would allow a single model to share information about multiple labels instead of training multiple models to learn each individually in a one vs rest paradigm. Could save training time and potentially improve performance on closely related multilabel tasks.

Additional Details

This will probably need to be supported on a per-model basis (via a new type of task? "TrainMultilabel"?), since it will be relatively easy to implement for ex. Transformer/MT-DNN but very difficult for official BERT.