Saddle Flow Dynamics: Convergence, Algorithms, and Application to Constrained Reinforcement Learning