What Matters In On-Policy Reinforcement Learning? A Large-Scale Empirical Study (arXiv’20)