PEARL: Preconditioner Enhancement through Actor-critic Reinforcement Learning