Investigating a Multi-Modal Attention-Based Deep-Learning Framework for Long-Term IMF $B_{z}$ Forecasting Based on CME Kinematics and Solar Magnetogram Data