Using cluster-robust standard errors when analyzing group-randomized trials with few clusters