Learning Abstract World Models with a Group-Structured Latent Space