Towards Efficient Diffusion-Based Image Editing with Instant Attention Masks