D^2ETR: Decoder-Only DETR with Computationally Efficient Cross-Scale Attention