More consistency in file formats

[dafnevk]

I plan to make the following changes:

• separate fields in Token for bracket corefs and head-based corefs
• Add a check that preprocessing took place before running code that assumes this
• Save the transformed dependencies in a different fields, so that the original tree is also stored
• use different file extension for conll-like files that have transformed tree and head-based mentions
• Make transform_tree method of sentence / dataset classes
• make preprocess_sentence part of sentence class
• make postprocess sentence part of sentence class
• add tests to check everything still works