Shortened LLaMA: A Simple Depth Pruning for Large Language Models