Different numbers of timepoints can matter because the connectivity estimates from longer series are usually more stable. I would not fix that by upsampling. A better approach is to compute connectivity from matched portions of the task across all subjects (for example the same blocks or the first common segment), convert correlations to Fisher z, and then compare groups at the second level. You can include the number of usable frames as a covariate only as a secondary sensitivity analysis, but not as the main solution if scan length is essentially confounded with group.