Optimize SurfaceLight sphere sample distribution

[snogglethorpe]

Currently Sphere::Sampler and Sphere2::Sampler just evenly sample over the half-sphere for the Surface::Sampler::sample_from_viewpoint method, which doesn't yield particularly good results.

It might be better to sample evenly over the field-of-view cone instead, e.g. with cone_sample.

When the sphere is far away, the sampling domain won't be particularly different, but the distribution will probably be better.

This affects issue #26.