Enhancing Convergence of Decentralized Gradient Tracking under the KL Property