LLM-based Fusion of Multi-modal Features for Commercial Memorability Prediction