Get Back Here: Robust Imitation by Return-to-Distribution Planning