OmniMAE: Single Model Masked Pretraining on Images and Videos