Provably Efficient Reinforcement Learning with Linear Function Approximation