[docs] document how to support operators with several vector arguments

[odow]

Closes #3576

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Comparison is base (283a99d) 98.38% compared to head (65ac069) 98.38%. Report is 2 commits behind head on master.

Additional details and impacted files

```diff @@ Coverage Diff @@ ## master #3577 +/- ## ======================================= Coverage 98.38% 98.38% ======================================= Files 37 38 +1 Lines 5629 5648 +19 ======================================= + Hits 5538 5557 +19 Misses 91 91 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[odow]

@blegat is this what you wanted:

https://jump.dev/JuMP.jl/previews/PR3577/manual/nonlinear/#User-defined-operators-with-vector-inputs

[blegat]

It might be more helpful to show how to deal with functions that cannot be traced and to show how to also get the derivative info with AD and pass it to @operator

[odow]

Is this part not sufficient? https://jump.dev/JuMP.jl/stable/manual/nonlinear/#Multivariate-functions

[blegat]

It doesn't say that you can use AD

[odow]

I'd rather not encourage people. We can already use ForwardDiff automatically. If anyone has a user-defined operator and wants to use a custom AD that isn't ForwardDiff, then I expect them to be very advanced users who can figure it out based on the signature of grad_f(g::AbstractVector{T}, x::T...) where {T}.

If we tell people to use Enzyme/Zygote/ReverseDiff.jl then we'll have to deal with the questions and quirks.

[odow]

This is good enough for now. A question of whether to add a third-party AD example is best left for a different PR. I'll make a note in the issue.