Nov fixes and updates

[iross]

• Tesseract4 has an issue where a word's bbox can take up the whole page (Noise characters recognized with bbox as the entire page tesseract-ocr/tesseract#1192) -- blackstack now skips those words
• Added simple wrapper script + envvar toggle to run docker-compose setup in either classified or training mode
• There was some potential weirdness where the list of labels was getting read from annotated docs instead of from the labels table. It would crash if you start a new model but don't have examples of each layer.
• A few python3 fixes -- 5dcda56 is a critical one. filter() in python3 is a generator, so len(filter(...)) was throwing an exception that was getting silently caught. As a result, things downstream thought all areas contained zero words and document- and area-level heuristics broke.
• Added annotated page dump (outputs the pages with areas labeled by category

[iross]

@jczaplew Want to give it a look + run-through and let me know if there's anything else you want clarified/fixed?