SuperLoRA: Parameter-Efficient Unified Adaptation of Multi-Layer Attention Modules