Added EddyStopwatch class for simple timing tasks

Branch intended to look at/fix problem with strange, single voxel, "holes" in the CPU predictions. It turns out it is caused by treating zeros in input data as NaN. It only affects the CPU version on the first iteration when input images are of integer type and when no topup field is supplied. Hence it will have minimal impact on the end results and fixing it would require quite large changes. I have therefore decided to drop it for now. As part of my explorations I discovered that the CPU parallelisation of the GP prediction is incredibly inefficient, and should be fixed. I also implemented a simple stopwatch class for use in basic benchmarking.