Blending MPC & Value Function Approximation for Efficient Reinforcement Learning