Beyond In-Distribution Performance: A Cross-Dataset Study of Trajectory Prediction Robustness