Resource-Efficient and Delay-Aware Federated Learning Design under Edge Heterogeneity