Global Reinforcement Learning: Beyond Linear and Convex Rewards via Submodular Semi-gradient Methods and Geometric Active Exploration in Markov Decision Processes: the Benefit of Abstraction accepted at ICML 2024!