Add doc embedding implementation during inference

[valedica]

This closes #338. The implementation follows the one of the method get_document_embedding from neuralcoref.train.document, which is the method that calculates document embeddings during training.