Optimal Batch-Size Control for Low-Latency Federated Learning with Device Heterogeneity