Safe reinforcement learning framework for high-obstacle climbing in heavy wheeled-legged robots