HiMAP: Learning Heuristics-Informed Policies for Large-Scale Multi-Agent Pathfinding