The random element of running GPS is taken out by using precomputed gradient orientations.
The old implementation set a random subset of bvecs to b=0 (or to b=2000).
This is now fixed ensuring that the full bvecs are used for each shell.

there are sevaral definitions of MK, most of which are not compatible with the simplified kurtosis model implemented by --kurtdir. So, I decided not to compute anything at all.

single- and multi-shell
exact fit (no noise)