The Super Learner library shared by the propensity (g) and outcome (Q)
nuisance regressions. Sharing one library is what makes the estimators in
dr_panel() comparable: any difference between them is then estimator-driven,
not implementation-driven.