Weigh Your Own Words: Improving Hate Speech Counter Narrative Generation via Attention Regularization