`let` scope causes nested differentiation error

As reported by @tkoolen here. Note from my comment there - this does not seem to be a case of pertubation confusion, but rather an an inference problem of some sort (or possibly that ForwardDiff is incorrectly relying on inference behavior in some way).

The bug can be triggered with:

using ForwardDiff, Base.Test
@noinline f83a(z, x) = x[1]
z83a = ([(1, (2), [(3, (4, 5, [1, 2, (3, (4, 5), [5])]), (5))])])
let z = z83a
    g = x -> f83a(z, x)
    h = x -> g(x)
    @test ForwardDiff.hessian(h, [1.]) == zeros(1, 1)
end

If the let block is removed, then everything works normally. According to @tkoolen, this is coincidentally is fixed by #247, but it would be very valuable for us to figure out why.

ref https://github.com/tkoolen/RigidBodyDynamics.jl/issues/347
cc @goretkin

Start by running the Julia reproduction in the issue, comparing the nested let version with the version where let is removed. Read the discussion in #247 and the linked RigidBodyDynamics.jl issue to investigate the inference behavior; done means explaining the scope-dependent failure and resolving the Hessian test without relying on the coincidental #247 fix.