Much better memory management + word-level training support

• Switched to a fit_generator implementation of generating sequences for training, instead of loading all sequences into memory. This will allow training large text files (10MB+) without requiring ridiculous amounts of RAM.
• Better word_level support:
  • The model will only keep max_words words and discard the rest.
  • The model will not train to predict words not in the vocabulary
  • All punctuation (including smart quotes) are their own token.
  • When generating, newlines/tabs have surrounding whitespace stripped. (this is not the case for other punctuation as there are too many rules around that)
• Training on single text no longer uses meta tokens to indicate the start/end of the text and does not use them when generating, which results in slightly better output.