A cube should only have a coord-system through dim-coords

[pp-mo]

Currently, a cube cube will have a coord-system if any of its coords do.
It is the first found, looking at dim-coords first.

But I think that is really a mistake.
When we support the extended grid-mapping syntax, it will become possible for auxcoords to have a coord system, but this would not be the "coord-system of the cube".

So I think the cube coord-system "of a cube" should be that of its dim-coords only, or None.

For example, if the cube is on a projection grid, it will typically also have 2d "true" lats+lons as aux-coords.
If the dim-coords don't have a coord system (for some reason), the cube should not report that it is "on" a latlon grid, as that would be definitely wrong.

Relates: #3388