Add optimizations for maximum entropy tokenization

Depending on the version of opennlp and whichever fork you use off of it (language of choice), you can extend tokenization to support certain pre-optimizations such as: preserving hashtags, urls, @mentions, email addresses, emoticons..etc. It would be nice if chalk supported these kind of extensive features.
