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.

data for evaluation

1), and to handle modification of existing image

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.

Tests both kurtosis flags (also --wls with --kurt flag)

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

add initial round of bed unit tests

See merge request !3