GH-3655: Tokenization on predict

[alanakbik]

This PR changes the way tokenization works during prediction. It is now possible for models to remember the specific tokenizer they were trained with. This allows us to ensure that when predicting tags, both the model and the sentence objects follow the same tokenization scheme. Theoretically, this should yield better prediction accuracy.

Specifically:

• Models that inherit from DefaultClassifier - if trained with a specific tokenizer - now set this tokenizer to any Sentence object that is passed to them during prediction. Same for the SequenceTagger.
• The Sentence object now remembers the tokenization used to generate Tokens. There is a new setter that allows setting a different tokenizer for an already created Sentence. If a different tokenizer is set and tokens are requested, this triggers a retokenization with the new tokenization scheme.
• All Tokenizers are now serializable and have equality defined.

Closes #3655