Allow changing default cache for instantiate

[blegat]

The main use case for this is that once a solver defines a fast copy_to method with a MatrixOfConstraint model, then it can use it as cache directly to avoid having to store an extra cache. As instantiate is used by JuMP, this will be what's used by default in JuMP.

[odow]

This is really only useful for the cache inside a bridge, correct?

But this seems like you want JuMP to go with Option 2 as discussed here: jump-dev/Clp.jl#114 (comment).

This has some pretty big implications that I don't fully understand. Like it's now not possible to modify a cache after copy_to has been called.

[blegat]

Like it's now not possible to modify a cache after copy_to has been called.

Yes, it is breaking. However, this cache is mostly there because the bridges need the underlying model to support incremental loading of the model, not to support modification if the solver don't support it.

If the user want to be sure to be able to do modification MOI.instantiate(optimizer_constructor, with_bridge_type=T) is not the answer. Indeed, if the solver implement incremental loading but not modifications then no cache is added and it will fail when modifying.
So a solver-independent code can assume that MOI.instantiate(optimizer_constructor, with_bridge_type=T) supports incremental loading but not that it can support modifications after copy_to is called. This was already true before this PR.
To make incremental loading work without copy_to, we might need a few tweaks though such that calling MOIU.final_touch in CachingOptimizer before attempting to synchronize the optimizer.

[odow]

Before merging, I'd like to see this implemented by a solver.

[blegat]

I was doing it for Clp and then I got hit by issues which lead me to #1386 and #1387. I'll continue when these two are done.