SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics