GLUE/SuperGLUE Tasks Improvements

Created by: slowwavesleep

Features

This PR deals mainly with the implementation/reproduction of various tasks from GLUE and SuperGLUE benchmarks, as well as introducing the more recent version of HuggingFace's transformers to the library.

• Bump transformers version from 2.8.0 to 4.6.0
• Compatibility fixes to various tasks related to the transformers version update
• Add RTE and MNLI configs
• Add a new task type and a corresponding pipeline: Multiple Choice
• Add helper functions and config for the Choice of Plausible Alternatives task (COPA)
• Revise affected NER and SQuAD pipelines and configs to accommodate them to the transformers update
• Add new pre-trained model download paths where applicable
• Update model scores in the docs
• Experimental multi-GPU support