On Efficient Reinforcement Learning for Full-length Game of StarCraft II