Dual optimization for convex constrained objectives without the gradient-Lipschitz assumption