Add fake termination status to Gurobi (so it become compliant with MOI.MEMORY_LIMIT)

[henriquebecker91]

This PR intends to solve #397. The PR creates a generic mechanism that allows Gurobi Error Codes to interpreted as termination status by the Gurobi wrapper. For now, it is used only for transforming OUT_OF_MEMORY in MOI.MEMORY_LIMIT, but new conversions only need to add a new element to a global internal Dict. The mechanism consist in one extra field for Gurobi.Optimizer (called fake_termination_status), a new internal function _check_ret_GRBoptimize, a new internal global _FAKE_TERMINATION_STATUS, and changes to _raw_status, Optimizer (constructor), is_empty, empty!, and optimize!.

The code is well commented. I did not add tests. If tests are deemed necessary, then I would like that someone more familiar with exception testing inside Gurobi.jl gave me some pointers.

The current code give some errors:

but they are the same obtained in the master branch:

and probably should be fixed in other PR:

[odow]

Nice. Are there any other return codes we should catch?

[odow]

ret_GRBoptimize::Cint?

[odow]

Are you on a new release of Gurobi? Good to see they fixed it.

[henriquebecker91]

Are you on a new release of Gurobi? Good to see they fixed it.

Sorry, fixed what exactly? The Farkas dual? I am using Gurobi 9.1.1.

[odow]

There was a bug in Gurobi. But they fixed it which is why you now see "Got correct result"

[henriquebecker91]

About the changes, I have no problem with the purely stylistic ones, and will probably make a new commit soon to solve them. However, some of the changes of names and how conditionals are done are pointing could be looked at with a little more attention.

fake_termination_status -> ret_GRBoptimize::Cint

model.retGRBoptimize = GRBoptimize(model)
_check_ret_GRBoptimize(model)

_FAKE_TERMINATION_STATUS -> _ERROR_CODES

Its seems to me that, instead of making this field only for the error codes that may be interpreted as TerminationStatuses, the idea is always saving the error code of the GRBoptimize, and then, if the termination status is queried, we look at the Dict. This alone is ok, even more elegant than what I have initially thought (I mean, probably haskey may end up called multiple times but then the overhead is minimal). However, I think in this case the Dict really should not be called _ERROR_CODES that gives the impression that all Gurobi _ERROR_CODES are stored there. Maybe _ERROR_TO_STATUS instead? It is shorter and not so misleading.

[odow]

Maybe _ERROR_TO_STATUS instead?

Sure.

[henriquebecker91]

Made the changes, the tests continue to pass:

[odow]

Looks good.

[henriquebecker91]

Made the change, tests continue the same: