Improving SAT Solver Heuristics with Graph Networks and Reinforcement Learning