Learning to Provably Satisfy High Relative Degree Constraints for Black-Box Systems