[Enhancement request] Target the #1 SOTA for v5.0

According to the huggingface blog, neuralcoref is an adaptation of [Improving Coreference Resolution by Learning Entity-Level Distributed Representations](https://paperswithcode.com/paper/improving-coreference-resolution-by-learning), which has an accuracy of 65.29 on Ontonotes.
There **is** a market for making SOTA papers more available and easily useable. It's a real need. Spacy here is an excellent choice. However.. 65.29 is **mediocre**.
The state of the art in 2021 is 81% of accuracy AKA a 16 % increase.. This makes the difference between a cool demo and something that can begin to become used in the industry.
As a reminder, the SOTA of every NLP/NLU task is easily available and generally up to date on https://paperswithcode.com/sota/coreference-resolution-on-ontonotes