Towards trustworthy multi-modal motion prediction: Holistic evaluation and interpretability of outputs