Eliminating Cross-modal Conflicts in BEV Space for LiDAR-Camera 3D Object Detection