Gradient Transformation: Towards Efficient and Model-Agnostic Unlearning for Dynamic Graph Neural Networks