Optimizing graph neural network architectures for schizophrenia spectrum disorder prediction using evolutionary algorithms