Neural Lyapunov Function Approximation with Self-Supervised Reinforcement Learning