PolyGym: Polyhedral Optimizations as an Environment for Reinforcement Learning