Latent Mixture of Symmetries for Sample-Efficient Dynamic Learning