Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning.