Towards Better Generalization for Neural Network-Based SAT Solvers