Improved Regret Bounds for Bandit Combinatorial Optimization