Understanding Gradient Descent on Edge of Stability in Deep Learning