Direct Preference Optimization for Neural Machine Translation with Minimum Bayes Risk Decoding