…nto Trogluddite-master

Trogluddite master

cache start words separately to avoid scans. avoid storing useless grams

- read brain from either yaml or plain text
- write learned corpus by appending to a plain text file
- use generators where possible to avoid big reads
- don't write new phrases if we don't learn from them
- get rid of <START> and <STOP> tokens, they can be implicit in the corpus

This should make us performant enough we can forgo multiprocessing and do everything in-process

merge