You're right that all_FA (and all_MD or all_ADC etc when you've run tbss_non_fa) may contain the values you need but if you want to look at the *skeleton* values that were actually used in the randomise calculation then you need all_FA_skeletonised (all_MD_skeletonised etc).
As you'll see if you open the files up in fslview they're 4D files with the 4th dimension being subject in (usually) alphabetical order. All_FA.nii.gz contains the FA maps warped into MNI space, and all_FA_skeletonised contains the skeleton values for each person in MNI space.
You could use some combination of fslmaths, fslsplit, fslstats or fslmeants to apply a mask and then get the mean values within those masks. Type each command without any arguments to get their usage.
I think fslmeants is probably your easiest bet: