Quasi-Newton operators compatible with GPUs

[tmigot]

I am wondering how much work it would be to make QN operators compatible with GPUs (CuArray for instance)?

Typically, my use case would be something like this

using CUDA, NLPModels, NLPModelsModifiers, NLPModelsTest
V = CuArray{Float64}
nlp = NLSLC(V)

CUDA.allowscalar()
list_QN = [LBFGSModel, LSR1Model, DiagonalPSBModel, DiagonalAndreiModel, SpectralGradientModel]
lnlp = list[1](nlp)
x = nlp.meta.x0
v = copy(x)
Hv = similar(x)
hprod!(lnlp, x, v, Hv)

These model modifiers internally call the QN operators from LinearOperators.jl, e.g. op = LBFGSOperator(T, nlp.meta.nvar).