Add 3-Augment support to auto_augment.py, clean up weighted choice handling, and allow adjust per op prob via arg string

Author: rwightman