Optimal Rates for Vector-Valued Spectral Regularization Learning Algorithms