Refactor TopkPooling into SelectTopK

[wsad1]

Towards #6455.
Do not review before #7307 is merged.

[codecov]

Codecov Report

Merging #7308 (114b9a4) into master (c566f5c) will increase coverage by 0.11%.
The diff coverage is 97.43%.

❗ Current head 114b9a4 differs from pull request most recent head 1cbb5ef. Consider uploading reports for the commit 1cbb5ef to get more accurate results

@@            Coverage Diff             @@
##           master    #7308      +/-   ##
==========================================
+ Coverage   91.40%   91.52%   +0.11%     
==========================================
  Files         437      438       +1     
  Lines       24306    24337      +31     
==========================================
+ Hits        22218    22274      +56     
+ Misses       2088     2063      -25     

Impacted Files	Coverage Δ
torch_geometric/nn/pool/select/topk.py	96.82% <96.82%> (ø)
torch_geometric/nn/pool/asap.py	90.41% <100.00%> (ø)
torch_geometric/nn/pool/pan_pool.py	95.74% <100.00%> (+0.09%)	⬆️
torch_geometric/nn/pool/sag_pool.py	100.00% <100.00%> (ø)
torch_geometric/nn/pool/select/__init__.py	100.00% <100.00%> (+100.00%)	⬆️
torch_geometric/nn/pool/select/base.py	96.42% <100.00%> (+96.42%)	⬆️
torch_geometric/nn/pool/topk_pool.py	100.00% <100.00%> (ø)

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

[rusty1s]

Hi @wsad1, I modified this PR quite a bit. Hope the changes work for you:

• I moved attention score computation to SelectTopK - I think this belongs there.
• Added tests for SelectTopK
• I reverted the addition of TopKConnect - Can we add that separately with corresponding test cases? I would also be in favor of renaming that to something more general, e.g., FilterGraph? The name should make it clear that we only maintain the nodes present in select_output.node_index. It may also require changes since it should respect the assignment matrix coming from SelectOutput
• Can you add documentation of SelectTopK in a follow-up?