Sim-and-Real Co-Training: A Simple Recipe for Vision-Based Robotic Manipulation