With the large-scale commercial deployment of fourth-generation (4G) and fifth-generation (5G) networks, cellular-based localization has emerged as a cost-effective solution for Internet of Things (IoT) applications. However, existing methods face two critical challenges: input sparsity and dimensionality explosion due to the exponential growth of Base Stations (BSs) in wide-area scenarios, and underutilization of diverse cellular signal parameters. To address these issues, we propose a Heterogeneous Graph Neural Network (HGNN)-based framework for seamless large-scale cellular localization. First, we model cellular associations as a heterogeneous graph of User Equipments (UEs) and BSs, bypassing fixed-dimensional input constraints inherent to conventional input form. This graph also explicitly utilizes cell-type relationships and topology features to enhance structural expressiveness. Then a dynamic pruning strategy is introduced to remove spatial irrevelent nodes, reducing structural noise. Further, meta-paths with explicit semantics are designed to constrain information propagation with interpretable spatial-semantic rules. Building on this graph, we propose a Heterogeneous Multi-feature enhanced graph ATtention network (HMAT) to fully utilize diverse features in the graph across metapaths for accurate and robust coordinate regression. Additionally, we introduce a lightweight variant, L-HMAT, with fewer parameters and zero-shot localization capability in new regions. Extensive experiments on three real-world cellular network datasets demonstrate that our approach significantly outperforms current state-of-the-art methods on large-scale cellular data.