Fix convert(::VariableIndex, ::ScalarAffineFunction) with zero coefficients

[odow]

Showed up in #2172

MosekTools fails tests because of:

It's pretty reasonable that we should skip zero terms when converting.

Check that it fixes Mosek

• https://github.com/jump-dev/MathOptInterface.jl/actions/runs/4974798432

[odow]

Hmm. I don't think I have solver-tests setup correctly. MosekTools doesn't seem to have installed the branch of MOI.

[odow]

https://github.com/jump-dev/MathOptInterface.jl/actions/runs/4975115966

https://github.com/jump-dev/MathOptInterface.jl/actions/runs/4975191716

[odow]

Okay, this fixes that particular issue. There's an unrelated problem with NormCone: https://github.com/jump-dev/MathOptInterface.jl/actions/runs/4975191716/jobs/8902130828

[blegat]

I often wanted to to this but usually, this is a sign you should use convert_approx

[odow]

I could expect convert_approx to handle something like 0.5x + 0.5x. But skipping zeros here seems like a low bar. It still requires exactly one non-zero with a coefficient of 1.0.

[blegat]

Yes, what I meant is that you probably also want to use convert_approx in that bridge if a solver returns something like 1e-10 * x.