As mentioned in #1298, our current "gradient" implementation is essentially a directional derivative.

We should support the “full” 3D gradient for a scalar field defined on an arbitrary horizontally unstructured surface mesh on the sphere.