Learning to Reject Sequential Importance Steps for Continuous-Time Bayesian Networks