Preference-Based Gradient Estimation for ML-Based Approximate Combinatorial Optimization