Complementarity-driven Representation Learning for Multi-modal Knowledge Graph Completion