Test: Add Mooncake AD testing to conv layer test infrastructure

[Parvm1102]

This is related to this issue #640
I have implemented Mooncake AD testing to Conv layers (conv.jl).

• Modified the test_gradients function in the GraphNeuralNetworks/test/test_module.jl file to accomodate Mooncake AD if test_mooncake flag is true.
• I found out that DConv and ChebConv are not fully compatible with mooncake.
• EGNNConv was excluded because it was still TODO.
• CGConv and GMMConv excluded because of segmentation fault (commented out earlier).

[Parvm1102]

@CarloLucibello Is this looking okay? I can proceed to check other layers too.

[CarloLucibello]

why this invokelatest?

[Parvm1102]

It is related to the world age, the import Mooncake and TestItemRunner's eval have different world ages and throws error. It is prevented by invokelatest.

[CarloLucibello]

tests related to this PR are failing (GNN julia 1).

On julia <1.12 mooncake testing should be skipped. You can define a global flag in test/runtests.jl

const TEST_MOONCAKE = VERSION >= v"1.12" 

[CarloLucibello]

Also, instead of calling Mooncake directly, we can call Flux.gradient(AutoMooncake(), ...) or ``Flux.withgradient(AutoMooncake(), ...)`.

See the testing function https://github.com/FluxML/Flux.jl/blob/master/test/test_utils.jl#L82,
which we should try to replicate here (leaving out reactant for the time being).

[Parvm1102]

Also, instead of calling Mooncake directly, we can call Flux.gradient(AutoMooncake(), ...) or ``Flux.withgradient(AutoMooncake(), ...)`.

See the testing function https://github.com/FluxML/Flux.jl/blob/master/test/test_utils.jl#L82, which we should try to replicate here (leaving out reactant for the time being).

@CarloLucibello, thanks for the detailed review! For Flux.withgradient(AutoMooncake(), ...), the FluxMooncakeExt hardcodes Mooncake.Config(friendly_tangents=true) and ignores the config field passed via AutoMooncake. (https://github.com/FluxML/Flux.jl/blob/ce4b8a081aec37d5f8144044a5a7940f47210551/ext/FluxMooncakeExt.jl#L11)

With friendly_tangents=true, Mooncake tries to deep-copy the closure's captured variables (including GNNGraph), which fails because DataStore contains Dict{Symbol,Any}. I'm calling Mooncake's API directly with the default config as a workaround.

I will include the version check for julia 1.12.

[CarloLucibello]

With friendly_tangents=true, Mooncake tries to deep-copy the closure's captured variables (including GNNGraph), which fails because DataStore contains Dict{Symbol,Any}. I'm calling Mooncake's API directly with the default config as a workaround.

If you could provide an example, this should be reported to the Mooncake repo. Ok having the workaroud for the time being.

[CarloLucibello]

There are a couple of layers that fail tests. We should:

• avoid testing them on mooncake so that we let the tests pass and can merge this PR
• create a list with the broken layers to keep track of them (e.g. in check compatibility with Mooncake #640 )
• create MWE and open issues in Mooncake.jl