BERT and SpanBERT for Coreference Resolution

# 🌟 New model addition

## Model description

This is a recent approach for co-reference resolution using BERT, implemented from the papers [BERT for Coreference Resolution: Baselines and Analysis](https://arxiv.org/abs/1908.09091) and [SpanBERT: Improving Pre-training by Representing and Predicting Spans](https://arxiv.org/abs/1907.10529), which is the current state of the art on OntoNotes (79.6 F1). It uses tensorflow 1.14.0.

Reason why this is interesting is it achieves strong improvements on the OntoNotes (+3.9 F1) and GAP (+11.5 F1) benchmarks. Also, I think it would be a nice addition to huggingface library, as it has only the neuralcoref as the coreference resolution module.

## Open source status

* [x] the model implementation is available: (https://github.com/mandarjoshi90/coref)
* [x] the model weights are available: (https://github.com/facebookresearch/SpanBERT)
* [x] who are the authors: (@mandarjoshi90, @jkkummerfeld, @wenyudu)