Towards Geometry-Aware Pareto Set Learning for Neural Multi-Objective Combinatorial Optimization