A Loss Curvature Perspective on Training Instabilities of Deep Learning Models