2-pass copy

[blegat]

As introduced in #48 (comment), we can distinguish layers of a composition of MOI instances as 1-IO, 2-IO or ∞-IO.
The InstanceManager, plays a central role as it has 1-I and ∞-O. This allows to make 1-IO layers and ∞-IO layers to be used together as follows :

All 1-IO ( InstanceManager ( All ∞-IO ( Solver ) ) )

One question remains : What should we do if there are 2-IO intermediate layers ?

• Scenario 1 : There are none.
• Scenario 2 : Make the InstanceManager implement 2-I so that we can do

All 1-IO ( InstanceManager ( All 2-IO ( InstanceManager ( All ∞-IO ( Solver ) ) ) ) )

• Scenario 2 : Make all solvers implement 2-I so that we can do

All 1-IO ( InstanceManager ( All ∞-IO ( All 2-IO ( Solver ) ) ) )

• Scenario 3 : Make all 1-IO implement 2-IO so that we can do

All 1-IO and 2-IO ( InstanceManager ( All ∞-IO ( Solver ) ) )

[blegat]

In the current situation, a layer has 4 choices:

1. Implements 1-IO, in this case, it can only contain 1-IO layers and solver
2. Implements 2-IO and 1-IO, in this case, it can only contain both 2-IO and 1-IO layers and solver
3. Implements ∞-IO, 2-IO and 1-IO, in this case it can contain layers and solver of any IO.
4. Implements 1-IO and the InstanceManager API (resetsolver!, ...) by passing it along to the internal instance, in this case it can contain the InstanceManager which can contain himself anything since it has ∞ output.

I was initially tempted to use approach 4) for bridges but after the discussion in #65, it seems that approach 2) is preferable. This means that the MOIU instance is on his own for handling custom constraint types : jump-dev/JuMP.jl#1152. We should have something different than bridges for instance-without-optimize! that want to support arbitrary constraints.

[blegat]

Thinking about it more, making bridges implement approach 2) isn't a good idea. Making them support 2-IO requires adding caching and complicate each bridge a lot. Approach 4) seems better, it also has the advantage of making bridge usable for ∞-IO solvers.
This doesn't change the fact that we need #84. If someone want to use a solver that supports a set not defined in MOI. The nonexistence of bridges for constraints with this set shouldn't block him.