The "-warp" option for 3dfractionize has nothing at all to do with the much-later-written nonlinear warping programs in AFNI. You'll have to manually transform the ROI dataset back to individual subject space, and then 3dfractionize that.
I've almost forgotten the 3dfractionize program, which I wrote about 20 years ago. And I'd completely forgotten its "-warp" option.