Do we want to support generators? (comprehensions, sum, etc.)

[dlfivefifty]

Is it possible to auto-diff the following:

julia> function mynorm(x)
       sqrt(sum(x^2 for x in x))
       end
mynorm (generic function with 1 method)

julia> mynorm([1,2,3])
3.7416573867739413

Or more generally, differentiate code with comprehensions ([x^2 for x in x]), minimum, etc.

This is hard since Generator uses anonymous functions.

map and broadcast also have this issue when passed an anonymous function. It's possible there's a work around by making a type BroadcastedDualVector that turns all the relevant function calls into broadcasted versions.