Node to source edge mapping has ragged array; not GPU compatible

[BSchilperoort]

The EdgesAtNode struct contains a mapping from nodes to source edges. This is represented in a ragged array. For example;

mapping = Vector([
    Vector([1]),
    Vector([2, 3]),
    Vector([4, 5, 6]),
])

This nested vector structure is not compatible with (julia's/KernelAbstractions) GPU arrays*. If the array cannot be copied to GPU we cannot use it inside a kernel.
To make it GPU compatible, we can use a 2D array, with 0 as padding value (as julia has no 0 index);

mapping = Array([
    1 0 0
    2 3 0
    4 5 6
])

Seeing as this will be an array of size (nodes, <node with maximum number of source edges>) the array shouldn't become much bigger than its current size.

The sum_at function needs to be modified to accommodate this, by skipping any 0 values.

function sum_at(A, inds)
    v = zero(eltype(A))
    for i in inds
        if i != 0
            v += A[i]
        end
    end
    return v
end

* Technically CUDA and ROCm support ragged tensors, but this is not available in Julia's generic GPU packages.