MatrixOfConstraints and IndexMap

Document that solvers can rely on the fact that `MatrixOfConstraints` does not implement deletion of rows and columns and hence, the `OptimizerCache` can be done in a way such that the index map is identity.

When using `OrderedProductOfSets`, the solver will need to keep the product of set as a field after `copy_to` to handle the mapping of indices to ranges, such mapping is needed for conic solvers anyway.

For linear solvers, there is also the choice of using `MixOfScalarSets` for which the mapping is trivial so no need to keep the product of set as a field.

Since that would make many solvers use `identity_index_map`, it gives incentive to make this function cheaper. One option is to make `CleverDict` even more clever with a bool `is_identity` and only start using `vector` when the mapping starts being non-trivial. That would also benefit to solvers not using `identity_index_map` but for which the mapping is trivial most of the time.

See https://github.com/jump-dev/Cbc.jl/pull/189#discussion_r818105943

Start by reviewing the relationships among MatrixOfConstraints, OptimizerCache, OrderedProductOfSets, copy_to, identity_index_map, MixOfScalarSets, and CleverDict, then read the linked Cbc.jl discussion. Done means the solver index-map assumptions and the proposed handling of product sets and identity mappings are documented, with any implementation scope agreed.

Rédigé par le modèle d'indexation à partir du texte de l'issue.