Handle Upper Limits Better

[guillochon]

Currently we just use a half-Gaussian for upper limits, but we really should use a proper error function form for upper limits. Not remembering where to look this up at the moment, but essentially there ends up being a slight score penalty even if the upper limit is satisfied, just because of the uncertainty in where the upper limit is actually located.

[guillochon]

A better way of handling upper limits is available in this paper, but the version that works for GP requires an iterative scheme (probably expensive). May be the only way to do it right though.

[guillochon]

Superseded by #95.