Pre-Training for Robots: Offline RL Enables Learning New Tasks from a Handful of Trials