Nonlinear kernels, dominance, and envirotyping data increase the accuracy of genome-based prediction in multi-environment trials