Implement hessian functions with different minibatch

[lrsantos11]

As I mentioned in Zulip,

I would like to use FluxNLPModels.jl and I need to use second order information. According to Flux.jl, one can use again gradient to compute the hessian.

In fact, I want a possible different batch to compute the hessian in relation to the gradient.

I saw the in FluxNLPModels.jl the NLPModels functions grad and objgrad are implemented.

• Should we implement the functions related to the hessian?
• If so, use of Zygote (as Flux.jl indicates) is preferable?