GNNGraph constructor is type unstable

[oxinabox]

It seems like the GNNGraph constructor switches between where it is backed by a dense or sparse matrix depending on sparsity.
This kinda type instability is consider not so bad because it is handled by the small unions optimization.

However, when you map over these the resulting collectio becomes a Vector{Any}.
Which does not benifit from small unions optimization.

And I think it is very normal to map over these, when constructing graphs for all your training data.

Noticed in CellBH/DLkitty#13

[CarloLucibello]

this shouldn't happen. Could you provide an example?