Biology-informed neural networks learn nonlinear representations from omics data to improve genomic prediction and interpretability