Alternative normalization for `Normal`ly distributed parameters?

[glwagner]

@adelinehillier, should we subtract the mean and divide by variance? We're doing the converse now (subtracting variance and dividing by mean). It seems dividing by the mean has undesirable effects, like the variance in unconstrained space diverging as the constrained mean approaches zero. For example, we can't have unconstrained parameters with unit normal distributions using this method. But subtracting mean and dividing by variance is always valid, and implies that the unconstrained prior is unit Normal(0, 1).

[adelinehillier]

This is definitely the way it should be!

[glwagner]

Nice! I bumped the version, then I'll merge.