SpatioTemporal Focus for Skeleton-based Action Recognition