Disentangling Masked Autoencoders for Unsupervised Domain Generalization