Structured Pruning for Efficient Generative Pre-trained Language Models