Dissecting the Runtime Performance of the Training, Fine-tuning, and Inference of Large Language Models