Splitting prepare_dataset into preparing the base dataset, and the tokenized dataset. This will help us to have further control over caching and loading data, eventually removing the storage of base dataset.

Continuing cache minimizing in new repository. Please see https://github.com/huggingface/DataMeasurements for full history

:bug: filter_vocab -> filter_words

:bug: really make sure log_files/ exists

:tada: init