Support Automatic Differentiation

If we had AD, Accelerate would be ready for creating Deep Learning frameworks on top of it.

I've found a [very nice post](https://rufflewind.com/2016-12-30/reverse-mode-automatic-differentiation) explaining a way of implementing it in Rust and Python.

The thing is, if we implement it _inside_ Accelerate it would be great, as all programs could be passed through AD.

The source for the obsolete `rad` package created by Edward Kmett is quite simple, and might be trivial to implement. (http://hackage.haskell.org/package/rad-0.1.6.3/docs/src/Numeric-RAD.html)