You have 2 options for dealing with this type of spatial non-stationarity, inside the AFNI bubble-universe.
(1) Ignore it. At the group analysis level, it is not clear at all that the inhomogeneity in smoothness of the FMRI noise is important, relative to to all the other statistical "sins" that are committed along the way.
(2) Eradicate it. That is, blur the time series data, before analysis, in a way that eliminates this non-stationarity. The AFNI program 3dBlurToFWHM will do this, blurring more in areas with less smoothness and blurring less in areas where the noise is already smooth. At that point, you can then analyze the time series data (e.g., via 3dREMLfit) as before, and continue.
or (3) -- try both methods and see how much different the end results are -- at the group level, not in individual subjects.
I don't know of any statistical tool that can correctly compensate for non-stationary smoothness in clustering at the group level, considering also the fact that the amount and spatial distribution of this non-stationarity would differ between subjects. Perhaps some brute force resampling/permutation method could deal with this issue, but I don't have or know of any software for this purpose.