Reliable Gradient-free and Likelihood-free Prompt Tuning