Learning World Models with Identifiable Factorization