Achieving Zero Constraint Violation for Concave Utility Constrained Reinforcement Learning via Primal-Dual Approach