Orthogonality-Promoting Distance Metric Learning: Convex Relaxation and Theoretical Analysis