AD error with rat42

Even though the AD backend for the residual Jacobian is defined, it returns an empty Jacobian:
```julia
julia> rat42_nls = rat42(; use_nls = true)
ADNLSModel - Nonlinear least-squares model with automatic differentiation backend ADModelBackend{
ForwardDiffADGradient,
ForwardDiffADHvprod,
EmptyADbackend,
EmptyADbackend,
EmptyADbackend,
SparseADHessian,
EmptyADbackend,
ForwardDiffADHvprod,
ForwardDiffADJprod,
ForwardDiffADJtprod,
SparseADJacobian,
SparseADHessian,
}

julia> jac_structure(rat42_nls)
(Int64[], Int64[])

julia> x = [99.94710393753867, 1.6483148108416925, -12.429444473714828]

julia> jac(rat42_nls, x)
0×3 SparseMatrixCSC{Float64, Int64} with 0 stored entries
```

There is also an error with the gradient of the NLP model:
```julia
julia> rat42_model = rat42()
ADNLPModel - Model with automatic differentiation backend ADModelBackend{
ForwardDiffADGradient,
ForwardDiffADHvprod,
EmptyADbackend,
EmptyADbackend,
EmptyADbackend,
SparseADHessian,
EmptyADbackend,
}

julia> obj(rat42_model, x)
9111.7101

julia> grad(rat42_model, x)
3-element Vector{Float64}:
NaN
NaN
NaN
```

@amontoison @tmigot Any ideas?