Data Scaling Laws in Imitation Learning for Robotic Manipulation