Module-wise Adaptive Distillation for Multimodality Foundation Models