A Semiparametric Instrumented Difference-in-Differences Approach to Policy Learning