MSWA: Refining Local Attention with Multi-ScaleWindow Attention