Leveraging fine-mapping and multi-population training data to improve cross-population polygenic risk scores