DropMAE: Masked Autoencoders with Spatial-Attention Dropout for Tracking Tasks