Fast Rates for Bandit Optimization with Upper-Confidence Frank-Wolfe