Convex Is Back: Solving Belief MDPs With Convexity-Informed Deep Reinforcement Learning