Learning to Guide a Saturation-Based Theorem Prover