ReGLA: Refining Gated Linear Attention