Affine groupwise registration uses data parallelism by image for SMP computation.

If a large number of small images are registered, the algorithm scales extremely poorly.

A second level of parallelism should be added, most importantly over the components of the cost function gradient. That will require some major changes, however, because there is currently only temporary space for one reformatted image, and the "Evaluate()" function always works on the object's m_Data[] array.