Shape Matters: Understanding the Implicit Bias of the Noise Covariance