Improve additionnal vocabulary management

[n1t0]

Fix #302

• Extracted Additional vocabulary from Tokenizer
• Added normalized option, to handle cases where the tokens should be extracted from the normalized version of the text. For example, when adding a token like yesterday, we expect it to match in the input text Yesterday I saw a cat if we have a Lowercase Normalizer.
• Added missing tests

This brings a breaking change, where AddedToken now requires a second argument is_special_token: bool, that is used to set the default normalized behavior. Special tokens shouldn't be normalized, and just extracted as given, while normal tokens should be normalized by default. The normalized behavior can be manually set on any token to override the default.