Several issues with interp_points

[rhugonnet]

The Raster.interp_points function has several urgent issues:

1. The mode argument is never used, and actually defines the wrong thing (it is the order argument that defines the type of interpolation in map_coordinates, mode is for behaviour around the margin of the image),
2. The prefilter argument of map_coordinates runs by default: https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.map_coordinates.html#scipy-ndimage-map-coordinates. This spline pre-filter will actually expand NaNs for up to 20 PIXELS (tested with order=3 on this data: https://github.com/ICESat2-SlideRule/coregistration/blob/master/xdem_test.ipynb),
3. Scipy's map_coordinates is used for performance (much faster than interpn), but only works on grids of equal spacing, need to combine with interpn,
4. We need to clarify/merge behaviour with value_at_coords, see Improve docstrings of Raster.value_at_coords and Raster.interp_points, or merge the two methods. #395.