add extra vector of storage inside QN models

[dpo]

The objective of this change is to move the QN updates to a callback instead of having it performed in a generic manner by the solver. This will allow for customized QN updates.

[Copilot]

Pull request overview

This PR adds an extra vector field v to the LBFGSModel and LSR1Model structs to store gradient differences (yₖ = ∇fₖ₊₁ - ∇fₖ). This change lays the groundwork for moving quasi-Newton updates to callbacks instead of having them performed generically by the solver, enabling customized QN updates.

Changes:

• Added v::S field to LBFGSModel and LSR1Model structs for storing gradient differences
• Updated constructors to initialize the new vector field using similar(nlp.meta.x0)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Package name	latest	stable
FletcherPenaltySolver
JSOSolvers
JSOSuite
LLSModels
NLPModelsIpopt
NLPModelsKnitro
NLPModelsTest
PartiallySeparableNLPModels
Percival
QuadraticModels
RegularizedOptimization
RipQP

[tmigot]

Thanks @dpo !
Should we also specialize the push! for these two operators ?

NLPModelsModifiers.jl/src/quasi-newton.jl

Line 218 in 083636e

function Base.push!(nlp::QuasiNewtonModel, args...)

[dpo]

Do you mean that push!() would always push model.v? I think we would want to leave as much freedom as possible.

[tmigot]

Sure, let's keep it as it is now. The rest looks good to me, thanks!

[dpo]

Thanks @tmigot and @MaxenceGollier !