BERT model functionality

Adding BERT functionality, including WordPiece subword tokenizer (as well as other common sub word tokenizers), pretraining, finetuning and saving BERT models, loading and using pretrained BERT models.

• Interface into huggingface tokenizers library for training new tokenizers
• Interface into BERT models in transformers library. For each task class: initialization, saving and loading, training for associated tasks, extracting outputs for specific task, task prediction, unit tests and integration tests as needed
  • Base class for extracting features (aka hidden states) from any BERT task model.
  • BERT pretraining class
  • BERT masked language model class
  • BERT sequence classification class
  • BERT token classification class optional (for now):
  • BERT multiple choice class
  • BERT question answering class
  • BERT co-reference fine-tuning class
• Add custom dataset and collator functionality for training
• Generalize interface for any Transformer model
• Documentation
  • API docs
  • Use cases
  • Code snippet usage examples
  • Update readme
• Passing all tests
• Developer documentation