VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Priors