Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models