Co-Activation Graph Analysis of Safety-verified and Explainable Deep Reinforcement Learning Policies