Only 5\% Attention Is All You Need: Efficient Long-range Document-level Neural Machine Translation