Missing implementation of doc embeddings during inference

Document embeddings are not calculated during inference in [neuralcoref.pyx](https://github.com/huggingface/neuralcoref/blob/60338df6f9b0a44a6728b442193b7c66653b0731/neuralcoref/neuralcoref.pyx), but they are left at zeros.

https://github.com/huggingface/neuralcoref/blob/60338df6f9b0a44a6728b442193b7c66653b0731/neuralcoref/neuralcoref.pyx#L717

This causes a mismatch between inference and training input features (doc embeddings during training are correctly calculated in [document.py](https://github.com/huggingface/neuralcoref/blob/60338df6f9b0a44a6728b442193b7c66653b0731/neuralcoref/train/document.py#L534)). Is it a bug or is it intentional? There is a call to a method `get_average_embedding` as a comment but it does not exist.