Improved Regret for Bandit Convex Optimization with Delayed Feedback