Tackling Long-Range Dependencies in Dynamic Range Compression Modeling via Deep Learning