Improving Transformer World Models for Data-Efficient RL