Fixed bug that meant that derivative images were wrong on the first iteration.

I had a report that there were discernible differences between GPU and CPU results for a specific data set. I found and fixed a bug that meant that derivative images were wrong on the first iteration when not supplying a --topup field. In addition to that I changed the Levenberg strategy so that it now always resets the Levenberg Lambda for each volume after each iteration. It will be slower, but should reduce the risk of a local minimum. The results are much more similar between CPU and GPU now, but still not identical. I will keep the branch open and keep looking.