M3Video: Masked Motion Modeling for Self-Supervised Video Representation Learning