Double-Linear Thompson Sampling for Context-Attentive Bandits