Nearly-Optimal Bandit Learning in Stackelberg Games with Side Information