Fix heteroconv aggregation + tests

[svilupp]

Fixes #332

Proposal:

• adds unique() to _reduceby_node_t to construct a valid tuple
• unique operation is blocked from ChainRulesCore
• adds tests for HeteroGraphConv aggregation

Question:

• Are you okay with this implementation?
• Is there a cleaner way to avoid trying to take gradients over unique? (I didn't want to block the whole signature, because it's a fairly generic one)

[CarloLucibello]

Nice, seems the right thing to do! Also very well tested, thanks.

[CarloLucibello]

Is there a cleaner way to avoid trying to take gradients over unique? (I didn't want to block the whole signature, because it's a fairly generic one)

The current solution is fine and we can merge as it is. At some point unique will be fixed by JuliaDiff/ChainRules.jl#736

[svilupp]

I've checked the nightly error.
It's in the same file as I've changed ("test/layers/heteroconv.jl"), but it seems to be unrelated. It passes locally on 1.10 beta2 and 1.9.3.

Error comes from outside the tests in the gradient() call, probably because of the getindex access?

grad, dx = gradient((model, x) -> sum(model(g, x)[1]) + sum(model(g, x)[2].^2), model, x)

Error:

HeteroGraphConv: Error During Test at /home/runner/work/GraphNeuralNetworks.jl/GraphNeuralNetworks.jl/test/layers/heteroconv.jl:1
Got exception outside of a @test
Can't differentiate boundscheck expression $(Expr(:boundscheck)).
You might want to check the Zygote limitations documentation.
https://fluxml.ai/Zygote.jl/latest/limitations

[CarloLucibello]

Flux itself is failing on nightly so at the moment I'm not paying attention to it.