OCReP: An Optimally Conditioned Regularization for Pseudoinversion Based Neural Training