add an NLP test with NaN returned from the callbacks

[mlubin]

Ref jump-dev/Ipopt.jl#133

[shadiakiki1986]

I'm trying to tackle this issue by writing tests for it in my forks of MOI and Ipopt. My results don't match with my expectations, so I have a few questions. This is very draft-ish with a lot of coding malpractices like function and variable names as well as unnecessary code from copy-paste.

1. Is my approach of implementing tests in MOI and calling them from Ipopt in line with how you're building the packages?
2. My first test max_x_is_inf is implemented in MOI here, but it errors when called from Ipopt here with Attribute MathOptInterface.VariableName() is not supported by the model. What I was aiming for with this test is to set up an optimization that goes to infinity and see if it starts popping up NaN's on the way to Inf. I ended up re-implementing this idea in my next test nancb_test (details below) but without the MOIU.loadfromstring, which I was hoping would work and be a very simple and compact implementation of tests for this issue. I think I'm missing something very obvious about the syntax of the string model here.
3. The next test nancb_test is defined in MOI here and called in Ipopt here. It is a simple max x st x>=1. It's mostly copied from the NLP HS071 test with modifications for the simpler model here. The termination status that ends up here is MOI.NORM_LIMIT, with primal status = UNKNOWN_RESULT_STATUS. The objective value comes out at 4.94, but I get it that whatever number here is invalid given the unknown status flag. I don't understand why it doesn't stop at something much larger, but maybe this has to do with some configurations for Ipopt that I missed.
4. The test nancb_test2 is defined in MOI here and called in Ipopt here. It's the same idea as above, but the problem is max 1/x st x>=-1. The idea was to get an Inf while converging to 0. The termination status comes out as MOI.ITERATION_LIMIT (different than the previous test), with primal status as MOI.UNKNOWN_RESULT_STATUS as before. The objective value comes out at 0.06. I was expecting this to come out as much larger (closer to Inf).
5. The last test nancb_test3 is defined in MOI here and called in Ipopt here. Here I bounded the variable to 0.01 <= x <= 1 and manually set NaNs in the gradient, jacobian, and hessian. Despite my bounding, I still get INVALID_MODEL (docs about it were added recently in the readme).
6. Thanks for getting this far. Am I in the right direction on this issue? I'd completely understand if it turns out that resolving this is just a 3-liner versus my 500 LOC.