Neuro-Symbolic Reinforcement Learning with First-Order Logic