Leveraging Continuous Time to Understand Momentum When Training Diagonal Linear Networks