LP-based policies for restless bandits: necessary and sufficient conditions for (exponentially fast) asymptotic optimality