The accuracy of different strategies for building training sets for genomic predictions in segregating soybean populations