Uncertainty-Encoded Multi-Modal Fusion for Robust Object Detection in Autonomous Driving