Add PyTorch SparseTensor support for GCNConv and gcn_norm

[EdisonLeeeee]

For gcn_norm, a PyTorch SparseTensor is converted to torch_sparse.SparseTensor and then converted back. Not really like this but have no better solution.

[rusty1s]

Thanks for this PR. I think this gets really tricky to implement with TorchScript support. Sadly, the full tests do not pass in this PR but I also failed to make them work :(

[EdisonLeeeee]

Yeah, I have been blocked by this problem and also failed to find a solution these days. It seems we should drop the TorchScript support for PyTorch SparseTensor for now.

[rusty1s]

Yeah, we don't have support for this anyway yet. The challenge is to ensure that original TorchScript support is maintained :(

[EdisonLeeeee]

Noticed now. The problem seems related to is_torch_sparse_tensor - unable to distinguish between the Tensor(dense) and torch.sparse.Tensor when converting to torchscript :(

[codecov]

Codecov Report

Merging #6033 (f08b95b) into master (11e576b) will increase coverage by 0.00%.
The diff coverage is 100.00%.

❗ Current head f08b95b differs from pull request most recent head cd28763. Consider uploading reports for the commit cd28763 to get more accurate results

@@           Coverage Diff           @@
##           master    #6033   +/-   ##
=======================================
  Coverage   84.35%   84.36%           
=======================================
  Files         365      365           
  Lines       20500    20512   +12     
=======================================
+ Hits        17293    17305   +12     
  Misses       3207     3207           

Impacted Files	Coverage Δ
torch_geometric/nn/conv/gcn_conv.py	98.13% <100.00%> (+0.19%)	⬆️
torch_geometric/utils/softmax.py	100.00% <0.00%> (ø)
torch_geometric/nn/norm/layer_norm.py	100.00% <0.00%> (ø)
torch_geometric/nn/conv/rgcn_conv.py	95.39% <0.00%> (+0.09%)	⬆️
torch_geometric/nn/dense/linear.py	87.40% <0.00%> (+0.18%)	⬆️
torch_geometric/sampler/neighbor_sampler.py	67.05% <0.00%> (+0.25%)	⬆️

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[EdisonLeeeee]

@rusty1s I've made the tests passed with some changes on gcn_norm when dealing with torch.sparse.Tensor, it is somehow tricky though.

[rusty1s]

This is quite clever. Thank you!