Training Communication-Efficient Mixture-of-Experts Language Models with Layer Re-Configuration