Computationally Efficient Horizon-Free Reinforcement Learning for Linear Mixture MDPs