Drops the use of a real AFNI series of transforms. Instead, generate a number of large-yet-plausible transformations for head motion.

I have visually checked a few things:

• NiTransforms correctly generates the dataset with motion, and when realigining it with the originally created transforms, indeed the dataset does not show any motion (uncomment line writing out ground_truth.nii.gz to see). Meaning, NiTransforms seems to be a reliable tool for this test.
• The correction by our estimator is ATM not great, but we are just testing the integration of all pieces (see Integration and validation tests with b=0 only #40 for comments on fine-tuning).

Follow-up: #84.
Part-of: #40.