Near Instance Optimal Model Selection for Pure Exploration Linear Bandits