Parametrized container

[blegat]

Still very much a draft. I'll add more details later.
When the user uses the container syntax to create constraints, MOI receives each constraint independently so ExaModels.jl has to recover the repeated structure. We discussed during the JuMP-dev hackathon of a way to communicate the repeated structure through MOI more explicitly and this PR is a POC the resulting design.
As a byproduct, this also gives a solution for #1654 since the constraint will be expanded at the MOI level with a bridge from IteratedFunction in Zeros to Scalar?Function in EqualTo (we need a way to know the scalar function type but that should be doable)
We can also recover the pretty printing of containers that we had in JuMP v0.18

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.58%. Comparing base (701251a) to head (d19658a).
Report is 2 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #3890   +/-   ##
=======================================
  Coverage   99.58%   99.58%           
=======================================
  Files          43       43           
  Lines        6031     6031           
=======================================
  Hits         6006     6006           
  Misses         25       25           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mlubin]

Does this change the semantics of when and how often terms are evaluated?
What if you write something like:

@constraint(model, [i in [:a, :b]], x + d1[database_lookup(i)] >= d2[database_lookup(i)], container = ParametrizedArray)

[blegat]

Good question, this will not change the semantic, your subexpressions depending only on i but not any JuMP expressions could be any code (although this doesn't work yet but that's the idea). Once you start interacting with JuMP expressions, we record the value (so here the value of database_lookup(i)) obtained for each value of the iterator and this value has to be a Float64. At the MOI level, we pass the value obtained when we evaluated database_lookup(i) so MOI receives the Float64 value of each sub-expression depending on this iterator. I assume you don't have sub-expression depending on two iterators at the same time yet ^^

[blegat]

Moved to https://github.com/blegat/GeneratorOptInterface.jl as this is very experimental so it will be easier to experiment as a separate package