Alleviating Distribution Shift in Synthetic Data for Machine Translation Quality Estimation