Cancellation-Free Regret Bounds for Lagrangian Approaches in Constrained Markov Decision Processes