Efficient Training of Language Models to Fill in the Middle