Spacy-pytorch-transformers

* Fix serialization.
* Update `pytorch-transformers` to support [DistilBERT](https://medium.com/huggingface/distilbert-8cf3380435b5).
* Add pre-packed [DistilBERT model](https://github.com/explosion/spacy-models/releases/tag/en_pytt_distilbertbaseuncased_lg-2.1.0).

* Add out-of-the-box support for [RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta).
* Add pre-packaged [RoBERTa model](https://github.com/explosion/spacy-models/releases/tag/en_pytt_robertabase_lg-2.1.0).
* Update to `pytorch-transformers` v1.1.
* Fix serialization when model was saved from GPU.

* Add support for [GLUE benchmark tasks](examples/tasks).
* Support text-pair classification. The specifics of this are likely to change, but you can see [`run_glue.py`](examples/tasks/run_glue.py) for current usage.
* Improve reliability of tokenization and alignment.
* Add support for segment IDs to the `PyTT_Wrapper` class. These can now be passed in as a second column of the `RaggedArray` input. See the [`model_registry.get_word_pieces`](spacy_pytorch_transformers/model_registry.py) function for example usage.
* Set default maximum sequence length to 128.
* Fix bug that caused settings not to be passed into `PyTT_TextCategorizer` on model initialization.
* Fix serialization of XLNet model.

* Handle unaligned tokens in extension attributes.

> ⚠️ **This version requires downloading new models.**

* Fix issue 15: Fix serialization of config and make models load correctly offline.
* Improve accuracy of textcat by passing hyper-parameters correctly (Adam epsilon, L2).
* Support pooler output for BERT model.
* Add `fine_tune_pooler_output` model architecture option for `pytt_textcat`.
* Add Glue benchmark script in [`examples/tasks/run_glue.py`](examples/tasks/run_glue.py).
* Improve overall stability.