Implement VV10 non-local correlation

[susilehtola]

The Vydrov-van Voorhis non-local correlation model is used in many density functionals. Since it is evaluated from the density on a quadrature grid and does not depend on the basis set, the logical place to put it would be in GauXC.

$$ E_\text{nlc} = \frac 1 2 \int n({\bf r}) K({\bf r}, {\bf r}') n({\bf r}') {\rm d}^3r {\rm d}^3r' $$

Also other types of non-local correlation models might be implemented in GauXC. On the solid state side, libvdwxc provides implementations for many non-local correlation functionals but their implementations assume a uniform grid like is used in plane-wave and pseudo-atomic orbital approaches.

[wavefunction91]

Paper for future reference
Vydrov and Van Voorhis, J. Chem. Phys. 133, 244103 (2010) [doi:10.1063/1.3521275]

[susilehtola]

It would be good to have both energies and gradients. VV10 is often evaluated non-self-consistently for single points, since the relaxation effect is small. The VV10 evaluation requires a 6D integral. Key to the efficiency of the approach is that both integrals are done on the same quadrature grid; this is not really well-described in the papers, and the typical - and erroneous - assumption is that the inner integral is done on the small grid and the outer is done on the XC grid. However, this turns out not to be necessary and the small grid is used for both the inner and outer integral.