Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders