On the difficulty of learning and predicting the long-term dynamics of bouncing objects