Is there a scalar equivalent of `value_gradient_and_hessian` for functions that accept a `Number` rather than an `AbstractArray`? That seems an asymmetric omission from the interface.