Learning to Solve Weighted Maximum Satisfiability with a Co-Training Architecture