Geometry of Neural Reinforcement Learning in Continuous State and Action Spaces