How to compute tensor

Hi, I have defined a tensor like presented in the documentation:

function tensor(f, x)
    n = length(x)
    out = ForwardDiff.jacobian(y -> ForwardDiff.hessian(f, y), x)
    return reshape(out, n, n, n)
end

But when defined like this the tensors are just full of NaNs...
I used a simple Rosenbrock function as an example.
f(x) = 100*(x[2]-x[1]^2)^2+(x[1]-1)^2
The gradient are hessians are working just fine.

Reproduce the reported tensor function in Julia with the Rosenbrock function and inspect the ForwardDiff.jacobian/ForwardDiff.hessian composition shown in the issue. Compare it with the tensor example in the documentation; done means determining whether the NaNs are expected behavior or a bug and recording the explanation or required documentation change.