Normalize 3D coordinates in interp_weights to improve Delaunay triangulation performance

## Problem
`interp_weights` can be slow for 3D interpolation for grids with large aspect ratios (x/y spans of tens of kilometers vs. z spans of tens to hundreds of meters). This flat point cloud is difficult for `scipy.spatial.Delaunay` to tessellate efficiently.

## Fix
Scale the source and destination coordinates so each dimension has the same range before building the tessellation. This can follow the same approach as `scipy.interpolate.griddata` with `rescale=True`. The interpolation results would be unchanged because the same scaling is applied to both source and destination points.