Fix StackOverflowError coming from MOI.Utilities.operate

[bachdavi]

No description provided.

[chriscoey]

as discussed here - let's attempt to fix this upstream in MOI rather than adding to the complexity of SolverAPI

[chriscoey]

The MOI fix jump-dev/MathOptInterface.jl#2285 unblocks us for the test case added here, so lets get the test case in after MOI is tagged and leave this PR up here as DNMY. Not super high priority but when we become more concerned with performance, I think we could attempt an even larger change here that builds the SAF/SQF/SNF directly bottom-up from the JSON3 data structure, rather than going through the SNF intermediate. Because right now we do JSON->SNF with json_to_snf then SNF->SAF/SQF/SNF (whichever most canonical form we can convert to) with nl_to_aff_or_quad.

[bachdavi]

@chriscoey Sure we can wait for the upstream fix. I'm leaving this PR open in case we want to improve the performance of nl_aff_or_quad. I've taken the large-model from another one of your PRs and the stack-based version is around 3x faster and consumes 2x less memory (both running with the MOI fix). It makes a lot of sense to try to directly go from JSON to the respective function instead of this intermediate.

[chriscoey]

Thanks @bachdavi. Let's keep working on this PR to try to extend it to go directly from JSON. Hopefully that will improve the efficiency even further.

[chriscoey]

Since the upstream MOI fix is controversial, I decided to get this PR in rather than wait. We can further improve the performance in future PRs.

I split out the JSON-to-MOI functions into a separate file to improve clarity.