Support edge features in GATConv

[dongkwan-kim]

Related to #810

Some design choices that I made:

• I thought that filling the scalar value for edge_attr of self-loops might not be a good option. Instead, I wrote code that merges the features of edges that include the node of the self-loop by scatter/reduce operations.
  • https://github.com/dongkwan-kim/pytorch_geometric/blob/master/torch_geometric/nn/conv/gat_conv.py#L74-L80
  • https://github.com/dongkwan-kim/pytorch_geometric/blob/master/torch_geometric/utils/loop.py#L77-L82
• It uses the source nodes of edges in merging their features. It is not relevant to use the source or target nodes for undirected graphs, but not for directed graphs. Do you think should we create args for choosing the source or target?
  • https://github.com/dongkwan-kim/pytorch_geometric/blob/master/torch_geometric/utils/loop.py#L101
• I could not find this kind of scatter/reduce operations for SparseTensor. So I put some assert lines for this.
  • https://github.com/dongkwan-kim/pytorch_geometric/blob/master/torch_geometric/nn/conv/gat_conv.py#L212-L216

[codecov-commenter]

Codecov Report

Merging #3078 (7444821) into master (fdc7d5a) will increase coverage by 0.06%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #3078      +/-   ##
==========================================
+ Coverage   72.84%   72.91%   +0.06%     
==========================================
  Files         282      282              
  Lines       13629    13662      +33     
==========================================
+ Hits         9928     9961      +33     
  Misses       3701     3701              

Impacted Files	Coverage Δ
torch_geometric/nn/conv/gat_conv.py	100.00% <100.00%> (ø)
torch_geometric/utils/loop.py	100.00% <100.00%> (ø)

---

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 fdc7d5a...7444821. Read the comment docs.

[rusty1s]

Finally had some time to look into this. This is great, thank you!