Add an option to pick operator norm or Frobenius to compute relative error.

Calculate the norm directly with singular values instead of computing the approximate of M and then calculate the norm with it.