Learning to See before Learning to Act: Visual Pre-training for Manipulation