There’s a little bit in the documentation here and here about this. Your voxel size is a little smaller than the current recommendation, and the fixel count is correspondingly larger, the latter of which strongly influences the memory requirements. The tracking parameters used for tractography on the group average template can also influence the total amount of fixel-fixel connectivity and hence the amount of memory required.
There’s a large impending update to MRtrix3 statistical inference code here. One small but highly consequential piece of these changes is a substantial decrease in the amount of RAM required by
fixelcfestats. You could try that, though I don’t know if the savings would be enough to successfully execute with your current data. Also note that the RAM reduction is not the only change in that code; there are a lot of other changes to the GLM and surrounding code, which you would need to read and be aware of if you were to utilise the results from such.