API for derivatives

[andyferris]

transform_deriv() was always rather clunky. Some way of enabling automatic differentiation while at the same time enabling manually written derivatives would be great. C.f. JuliaDiff/ForwardDiff.jl#89

Not sure how the API could work out. Something like

tform(x; deriv = true)

is certainly possible, but maybe not ideal.

[c42f]

I think we should be overloading call to take something like ForwardDiff.Dual here (since it passes a point along with a set of partials), and keep it all single-argument to make chaining as transparent as possible.

Needs some reading I guess to understand how ForwardDiff is currently doing things and whether we can define some functions which don't force us to depend explicitly on ForwardDiff. (Hmm, and what about reverse diff? I guess the components you need there are basically just ForwardDiff on the inv() of a transformation? That makes me wonder whether defining inv() for the math functions in Base would be a powerful tool for building reverse diff tools... only in 0.5 of course :-) )