Improving Token-Based World Models with Parallel Observation Prediction