Make R2N more flexible for users

[MaxenceGollier]

I have two suggestions for R2N that would be very useful for me. I will try to implement them in a near future.

1. When I use R2N with the full Hessian, I'd need the Hessian stored in a sparse matrix format and to be able to access it through R2NModel I suggest having two different R2NModel structs: one called R2NModel_op that has the current implementation, i.e the one with the linear operator for the Hessian stored:
   

   RegularizedOptimization.jl/src/R2NModel.jl

   Lines 17 to 25 in 9bb4a9a

   	mutable struct R2NModel{T <: Real, V <: AbstractVector{T}, G <: AbstractLinearOperator{T}} <:
   	AbstractNLPModel{T, V}
   	B::G
   	∇f::V
   	v::V
   	σ::T
   	meta::NLPModelMeta{T, V}
   	counters::Counters
   	end

   
   the other one would be called R2NModel_full and would have the Hessian stored in, say, SparseMatrixCOO format. The type of R2NModel used by R2N would be dispatched as a keyword argument in the R2NSolver constructor (defaulting to the LinearOperator version). For this to work out we'd also need to modify the line

   RegularizedOptimization.jl/src/R2N.jl

   Line 420 in 9bb4a9a

   solver.subpb.model.B = hess_op(nlp, xk)

   , replacing it with hess_coord when the full hessian option is used.

2. In a constrained optimization context, I'd need to make a different update for the quasi-Newton approximation, with the difference of gradients of Lagrangians instead of the objective. I think the following lines should be a function called update_qn! that can be modified by the user:
   

   RegularizedOptimization.jl/src/R2N.jl

   Lines 415 to 419 in 9bb4a9a

   	if quasiNewtTest
   	@. ∇fk⁻ = ∇fk - ∇fk⁻
   	push!(nlp, s, ∇fk⁻)
   	end

   .

Please let me know what you think @dpo @MohamedLaghdafHABIBOULLAH !

[MohamedLaghdafHABIBOULLAH]

Interesting suggestion!