Preventing Value Function Collapse in Ensemble Q-Learning by Maximizing Representation Diversity