Injecting Numerical Reasoning Skills into Language Models