Improving Inference for Neural Image Compression