Flexible and Effective Mixing of Large Language Models into a Mixture of Domain Experts