Environment Parameter Gradient Theorem for Policy-Environment Co-Design in Reinforcement Learning