The Oversight Game: Learning to Cooperatively Balance an AI Agent's Safety and Autonomy