Masked Autoencoding for Scalable and Generalizable Decision Making