dataloader support for vector of graphs

[CarloLucibello]

Batching graphs is faster than selecting graphs from a batched graph, so we should reconsider how we feed data to the DataLoader.

using Flux, GraphNeuralNetworks, BenchmarkTools

function test(g)
    loader = Flux.DataLoader(g, batchsize=100)
    a = 0 
    for d in loader
        a += 1
    end
    return a
end

n = 5000
s = 10
data = [rand_graph(s, s, ndata = rand(1, s)) for i in 1:n] 
x1 = Flux.batch(data) 
x2 = Flux.batch([rand(s + s + s + s) for i in 1:n]) #source+target+data+extra

@btime test($x1);   #  1.295 s (2502 allocations: 6.17 MiB)
@btime test($x2);   #  357.595 μs (152 allocations: 1.61 MiB)
@btime test($data); #  65.288 ms (227002 allocations: 27.00 MiB) # this PR

Close #141.

[codecov]

Codecov Report

Merging #143 (9b59000) into master (f935e8d) will increase coverage by 0.16%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #143      +/-   ##
==========================================
+ Coverage   85.38%   85.55%   +0.16%     
==========================================
  Files          15       15              
  Lines        1259     1253       -6     
==========================================
- Hits         1075     1072       -3     
+ Misses        184      181       -3     

Impacted Files	Coverage Δ
src/GNNGraphs/gnngraph.jl	80.00% <100.00%> (+5.71%)	⬆️
src/GNNGraphs/utils.jl	74.00% <0.00%> (-1.25%)	⬇️
src/GNNGraphs/convert.jl	90.08% <0.00%> (-0.32%)	⬇️
src/GNNGraphs/query.jl	92.98% <0.00%> (-0.05%)	⬇️
src/GNNGraphs/transform.jl	96.71% <0.00%> (-0.04%)	⬇️
src/GNNGraphs/sampling.jl	100.00% <0.00%> (ø)
src/layers/conv.jl	79.52% <0.00%> (+0.17%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f935e8d...9b59000. Read the comment docs.