Augmented Lagrangian Multiplier Network for State-wise Safety in Reinforcement Learning