Changes to how derivatives are calculated to ensure that CPU and GPU results are more similar

This is a continuation of my investigation into why CPU and GPU results were notably different in a dataset given to me by Wenchuan and Xinyu. There was a problem when calculating the derivatives on the first iteration when running eddy without a topup fieldmap on the GPU. In addition, there was a different problem with how the derivatives were calculated on the GPU. That was related to using an inverse of a field to guide the inversion of another very similar (minimally perturbed) field. That problem had a much smaller, but still non-negligible, effect. These problems have both been fixed.