Next-Latent Prediction Transformers Learn Compact World Models