Allow Reflections by Removing Positivity Constraint for `LocationScale`

[ParadaCarleton]

Many asymmetric distributions are more easily expressed with a location-scale transformation that has a negative scale parameter. For instance, the lower tail of a variable can be modeled as a Generalized Pareto distribution with a negative scale. I propose removing the positivity constraint on LocationScale.σ to allow this.

[devmotion]

A quick summary of my comments on Slack: I don't think LocationScale should support it since it would not be a proper scale anymore and not correspond to the common definitions of location-scale families. I'm not opposed to a differently named type (and possibly deprecating LocationScale) that supports reflections.