More robust transformations and scaling adjustments

[fonnesbeck]

This replaces #1430, which was merged by accident and reverted.

Though this does not directly solve #1428, it does make transformations more robust, and puts better default bounds (more conservative) on the scaling guess when none is provided.

[twiecki]

I actually think this is required due to the sign disappearing when doing **-2

[springcoil]

This gets my vote to merge. @twiecki what do you think?

[twiecki]

I think this looks good. The new scaling min/max makes a bit weary though that we might be breaking models that have a huge scale on some values.

[fonnesbeck]

That's why I included the scaling_bound argument -- so that it can be overridden. Certainly in the case of #1428 a zero was passed as tau, and 1e-10 was too small as a substitute. When it comes down to guessing the scaling, we should perhaps be happy with something that works, rather than an "ideal" value. Having a conservative scaling (i.e. larger for small values and smaller for large values) shouldnt break the sampler, should it?

[twiecki]

Good point.