Sum rules (may?) have numerical errors for mesh interpolation orders > 3

[ceriottm]

The new test_derivatives in test_mesh_interpolator.py checks that the cell and atomic position derivatives of the mesh sum is zero, which should be exact given the sum rule in the interpolation. However, for high interpolation orders the derivatives are small (1e-5) but not zero. Maybe it's just numerics, maybe it's a bug...