Constrained Entropic Unlearning: A Primal-Dual Framework for Large Language Models