Learning 3D Representations from 2D Pre-trained Models via Image-to-Point Masked Autoencoders