Implement round() slightly different, so that for negative ndigits no additional errors happen in the last step. The trick is to avoid division by 0.1**n -- multiply by 10.0**n instead.