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