LLSModel variants

[abelsiqueira]

Currently LLSModel stores the COO of a matrix, and its constructors are either from the rows, cols, vals or A.
The variants would be:

• LLSModel that keeps the matrix A, whether it's sparse or dense. It should just use the regular matrix operations.
• LLSModel that keeps the COO of a matrix. It should use internal like coo_prod.
• LLSModel that keeps a LinearOperator.

There was a tentative here: JuliaSmoothOptimizers/NLPModels.jl#282