Refactor GenericNonlinearExpr and NonlinearOperator constructors

[odow]

Alternative for #3451

If we get rid of the "let's check if we can infer the variable type" then a lot of things become much easier.

[codecov]

Codecov Report

Patch coverage: 100.00% and project coverage change: -0.01% ⚠️

Comparison is base (a157126) 98.11% compared to head (cdecbb3) 98.11%.
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3489      +/-   ##
==========================================
- Coverage   98.11%   98.11%   -0.01%     
==========================================
  Files          37       37              
  Lines        5531     5529       -2     
==========================================
- Hits         5427     5425       -2     
  Misses        104      104              

Files Changed	Coverage Δ
src/nlp_expr.jl	99.51% <100.00%> (-0.01%)	⬇️
src/operators.jl	96.83% <100.00%> (-0.02%)	⬇️

☔ View full report in Codecov by Sentry.

📢 Have feedback on the report? Share it here.

[blegat]

I like this

[odow]

For future reference, here were the functions I removed

function variable_ref_type(
    ::Type{GenericNonlinearExpr},
    ::AbstractArray{T},
) where {T<:AbstractJuMPScalar}
    return variable_ref_type(T)
end

moi_function(x::AbstractArray) = moi_function.(x)

jump_function(model::GenericModel, x::AbstractArray) = jump_function.(model, x)

function test_nonlinear_operator_vector_args()
    model = Model()
    @variable(model, x[1:2, 1:2])
    op_det = NonlinearOperator(LinearAlgebra.det, :det)
    @inferred log(op_det(x))
    z = [2.0 1.0; 1.0 2.0]
    @inferred log(op_det(z))
    @objective(model, Min, log(op_det(x)))
    @test isequal_canonical(objective_function(model), log(op_det(x)))
    f = MOI.get(model, MOI.ObjectiveFunction{MOI.ScalarNonlinearFunction}())
    g = MOI.ScalarNonlinearFunction(:det, Any[index.(x)])
    h = MOI.ScalarNonlinearFunction(:log, Any[g])
    @test f ≈ h
    @test moi_function(log(op_det(x))) ≈ h
    return
end

[pulsipher]

I understand the motivation for this. Unfortunately, it will break InfiniteOpt, but I can look into refactoring the workflow. If supporting array arguments necessitates this change, then I support it.

[odow]

We can probably just keep this method then. It's not necessary. Let me make some changes.

[odow]

Try this

[odow]

I confirmed that this works with InfiniteOpt: infiniteopt/InfiniteOpt.jl#321

[pulsipher]

Thanks, you beat me to it :)