Rethinking the "Heatmap + Monte Carlo Tree Search" Paradigm for Solving Large Scale TSP