Problem-Complexity Adaptive Model Selection for Stochastic Linear Bandits