Learning Symmetric Embeddings for Equivariant World Models