GECToR 2.0.0

1. Write predictor class that can do multiple applications of forward() in model to replicate gector's iterative error correction.
2. Manually loading the model onto a cuda device should not be happening inside gec_model this is handled by allennlp's predictor base class
3. Modify pretrained gec model archive to use parameters for updated token embedder, token indexer, model etc.
4. Gector doesn't provide a model archive, they just provide the weights file. We should definitely make a model archive file.
5. Make model archive file for use with allennlp's Predictor.from_path() method and allennlp predict command
6. A config.json file will need to be written. These specify all the parameters used during training which we will need to extract from hardcoded values in Gec_model
7. Write unit test verifying that modified gec model archive can be used to do error correction on plaintext sentences.
8. Override json_to_instance behavior in predictor class so that gector's bespoke tokenization module is used.
9. Make the output from the predictor match the output from gec_model isn't accounted for.
   • Not correcting short sequences (<4 tokens),
   • Adding gector's start tokens to the input before correcting and stopping the iterations if no corrections were made for the previous iteration.