3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations