Variance-Dependent Regret Bounds for Non-stationary Linear Bandits