Text this: Sparse kernel models provide optimization of training set design for genomic prediction in multiyear wheat breeding data