Primal-Dual Continual Learning: Stability and Plasticity through Lagrange Multipliers