Federated Graph Learning (FGL) allows clients to collaboratively train Graph Neural Networks (GNNs) without exposing their private data. Nevertheless, FGL struggles with severe non-IID data issues, leading to degraded performance of the global model, especially across diverse domains. Our analysis reveals that high data heterogeneity in FGL likely arises from notable distribution differences between graph structures and features, which significantly complicate the simultaneous alignment of the global model. Motivated by this, we propose FedDense, an efficient FGL framework that decouples the learning and sharing of structural and feature information. To better acquire structural knowledge regardless of graph features, FedDense first explicitly encodes graph structures with a separate GNN channel. The structural channel is then shared among clients, while the feature learning remains locally, ensuring that the global model reconciles only the structural knowledge, thereby reducing heterogeneity in FGL. To further facilitate knowledge acquisition efficiency of both local features and shared structures, FedDense introduces a novel Dual-Densely Connected (DDC) architecture where each layer in the local feature channel is connected to all preceding layers from its own and the shared structural channel. Extensive experiments demonstrate that FedDense with narrow layers consistently outperforms baselines, achieving higher performance while minimizing resource costs.