Improved Regret Bounds for Online Kernel Selection under Bandit Feedback