Efficient Distance Approximation for Structured High-Dimensional Distributions via Learning