Make ForwardDiff conditional dependence

[mschauer]

Also give "definite" version of derivative helper functions such as

function negpartiali(f, d)
    id = collect(I(d))
    ith = [id[:,i] for i in 1:d]
    function (x, i, args...)
        sa = StructArray{Dual{}}((x, ith[i]))
        δ = -f(sa, args...).partials[]
        return δ
    end
end

avoiding perturbation confusion and https://discourse.julialang.org/t/gradient-and-directional-second-derivative-with-forwarddiff/63429