This contains more fixes for LabelGeometryMeasures.

The median was only added relatively recently, when switching out the deprecated ITK filter. But it is approximated and derived from a histogram, the default is to use 255 bins over the whole range of the image. This can get really imprecise when there are outliers or a large dynamic range, eg, t-statistic maps - this is how the problem was brought to my attention.

I experimented with using a mask of all the labels to define the histogram range for intensity voxels inside the mask. I also increased the number of histogram bins. It works better, but still, I'm not sure it's a good idea to include imprecise information, and I expect most users who really want median would rather compute it exactly with numpy or similar.

Mean, Min, Max, and Sum (called "IntegratedIntensity" for backwards compatibility) are computed exactly and remain in this version.