Add truncated Voronoi cells

[lcgraham]

Update certain Voronoi approximations to provide a cut-off distance (or directed distances) in a particular metric (thus defining a "cover" of any Voronoi cell by a simpler box defined by distance(s) and metrics). This can be used to speed up nearest neighbor searches since comparisons would not even need to be made with respect to samples that are farther away than the cut-off distances. It is very much like creating a pre-processing "filter" that is much like an indicator/characteristic function of the covering sets. This is particularly useful for emulation.

[eecsu]

Using gradients to define the scaled boundary box by
$$ (x - g)^\top diag(\abs{\nabla q}) (x-g) \leq \text{bd. tol.} $$

[lcgraham]

This truncated_voronoi_sample_set should be a subclass of voronoi_sample_set (see #178) with the additional attributes self._radii and self._normalized_radii whereself. _normalized_radii is the radii of the Voronoi cell when self._domain is normalized to the unit ball. If gradients are present these should be able to be set with self.set_normalized_radii() and self.set_radii(). Query will need to be updated to include distance_upper_bound as follows: self._kdtree.query(x, p=self.p_norm, distance_upper_bound=self._radii)

[lcgraham]

Note that these are not centroidal Voronoi tesselations meaning that the centroid is NOT the generator of the Voronoi cell. What we desire for the radius is actually :math:sup_{\lambda \in \mathcal{V}_{i, N}} d_v(\lambda, \lambda^{(i)})