TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation