How to Evaluate Machine Learning Approaches for Combinatorial Optimization: Application to the Travelling Salesman Problem