Toward Discretization-Consistent Closure Schemes for Large Eddy Simulation Using Reinforcement Learning