Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model