Conformal Symplectic Optimization for Stable Reinforcement Learning