Think before you speak: Training Language Models With Pause Tokens