Adding option to return info about sampled graph

[mszarma]

It's enabling the hierarchical tensor usage
and significant performance improvement

PyG part: pyg-team/pytorch_geometric#6661

[codecov-commenter]

Codecov Report

Merging #197 (1f0a3f0) into master (1760817) will increase coverage by 0.11%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #197      +/-   ##
==========================================
+ Coverage   83.37%   83.49%   +0.11%     
==========================================
  Files          26       26              
  Lines         842      848       +6     
==========================================
+ Hits          702      708       +6     
  Misses        140      140              

Impacted Files	Coverage Δ
pyg_lib/csrc/sampler/neighbor.cpp	100.00% <ø> (ø)
pyg_lib/csrc/sampler/cpu/neighbor_kernel.cpp	88.72% <100.00%> (+0.53%)	⬆️

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

[rusty1s]

I think code-wise this looks all good. Remaining questions would be

• Can we clean-up dispatching logic? Do you have any solution for this?
• Would prefer some renaming ala num_nodes_per_hop or similar
• I'm not totally sure why we gate behind disjoint in the heterogeneous case but not in the homogeneous one.

[mszarma]

I think code-wise this looks all good. Remaining questions would be

• Can we clean-up dispatching logic? Do you have any solution for this?
• Would prefer some renaming ala num_nodes_per_hop or similar
• I'm not totally sure why we gate behind disjoint in the heterogeneous case but not in the homogeneous one.

Thanks you @rusty1s for the review.

• Can we clean-up dispatching logic? Do you have any solution for this?
  I would recommend to put it to the PyG backlog and then think how to tackle that. Would prefer not to mix it. Move forward the feature part separately and then as separate task think how to refactor that dispatching logic. My gut feeling is that it may need a major refactor.
• Would prefer some renaming ala num_nodes_per_hop or similar
  That's a good idea. Done.
• I'm not totally sure why we gate behind disjoint in the heterogeneous case but not in the homogeneous one.
  The difference in hetero flow is that we are using the dict/Phmap. The issue is in usage of node_t: https://github.com/pyg-team/pyg-lib/blob/master/pyg_lib/csrc/sampler/cpu/neighbor_kernel.cpp#L220. In the disjoint hetero flow compiler point out that phmap cannot match/use the pair type. I didn't investigated usage of disjoint/temporal flow so my recommendation would be to enable feature firstly without disjoint support. WDYT @rusty1s ?