Leveraging Memory Copy Overlap for Efficient Sparse Matrix-Vector Multiplication on GPUs