Worst-Case Regret Bounds for Exploration via Randomized Value Functions