Optimizing over subsequences generates context-sensitive languages