Temper randomness in 3D GWN method

[jcs15c]

I've noticed that the text primal_solid_angle_test fails a bit more frequently than I'd like, as the underlying 3D GWN method has a few random components that fail with probability 0 mathematically, but probability >0 in floating point (by my estimate, about 0.1% of the time).

In particular, this PR addresses:

• Within the GWN algorithm, a re-cast ray was taken to be a random unit vector. This has been reworked to use a Halton sequence projected onto the unit sphere, which also has the advantage of more reliably distributing the cast attempts over the unit sphere.
• The per-surface cast-direction is taken to be the surface's average normal, plus a small perturbation. This has been changed so that the perturbation is seeded deterministically based on a (possibly too-simple) hashed function of the surface bounding box.

I've also added a simple test to verify this in primal_solid_angle.cpp.

[kennyweiss]

Thanks!
Would it make sense for halton to be a core utility function w/ unit tests?
Perhaps not this time, but the next time we need it (?)

Please update the RELEASE_NOTES about this change.

[jcs15c]

Thanks! Would it make sense for halton to be a core utility function w/ unit tests? Perhaps not this time, but the next time we need it (?)

It could get put it in axom::utilities alongside binomial_coefficient, but yes, since it's just used in this single lambda function for now, I think it'd be overkill. However, I do like Halton sequences on an aesthetic level, and will be eagerly looking for other places they could be useful.