MMRL++: Parameter-Efficient and Interaction-Aware Representation Learning for Vision-Language Models | Synapse