Decentralized Learning Strategies for Estimation Error Minimization with Graph Neural Networks