Scaling Laws for Uncertainty in Deep Learning