Feat/torch

Created by: dilyararimovna

• torch_trainer which is needed to model.train() and model.eval() settings. By default, torch models are set to train mode, so for build_model correct usage we still need to do eval in init of each torch model. We can add try model.eval everywhere in build_model, interact mode, etc but it looks very kostyl'no.
• base class TorchModel which saves, loads, initializes model on pytorch.
• torch_classification_model as an example of Neural Network implementation for classification task.
• BERT preprocessors on torch are almost all implemented
• BERT-based classification model on torch (with config example)
• BERT-based ranker model on torch (with config example)
• BERT-based ner model on torch (with config example)
• BERT-based SQuAD model on torch (with config example)
• BERT-based summarization model on torch (with config example)
• files with example models from configs are not all ready (will be uploaded to share later)