Leveraging Pre-Trained Models for Multimodal Class-Incremental Learning under Adaptive Fusion