A Riemannian Geometry Approach to Reduced and Discriminative Covariance Estimation in Brain Computer Interfaces