Add a fast Laplacian evaluator

[wavefunction91]

The mGGA implementation introduced in #84 evaluates the full hessian of the collocation matrix to evaluate the Laplacian. In principle, we should be able to hack gau2grid to produce the Laplacian directly (requiring less memory and fewer data transposes)

[wavefunction91]

This is partially addressed in #111 - GPU code directly evaluates the Laplacian rather than the Hessian