Faster thin singular value decomposition

Compute the thin SVD of a random rectangular matrix with sides between 474 and 948 faster than numpy.linalg.svd, with orthonormal factors.