Boosting Generalization in Diffusion-Based Neural Combinatorial Solver via Energy-guided Sampling