Balanced Gradient Penalty Improves Deep Long-Tailed Learning