Decision-Focused Learning with Directional Gradients