MAE-DFER: Efficient Masked Autoencoder for Self-supervised Dynamic Facial Expression Recognition