Interpretable Sparsification of Brain Graphs: Better Practices and Effective Designs for Graph Neural Networks