ConvMAE: Masked Convolution Meets Masked Autoencoders