Sampling according to max_sample within AddMetaPaths

[RexYing]

This functionality is used in cases where certain metapath edge types result in very dense adjacency.
If max_sample is set, in expectation this number of samples will be sampled per node for each edge index.

[codecov]

Codecov Report

Merging #4750 (b6546fd) into master (09f25e9) will decrease coverage by 1.89%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #4750      +/-   ##
==========================================
- Coverage   84.37%   82.48%   -1.90%     
==========================================
  Files         324      324              
  Lines       17304    17313       +9     
==========================================
- Hits        14601    14281     -320     
- Misses       2703     3032     +329     

Impacted Files	Coverage Δ
torch_geometric/transforms/add_metapaths.py	100.00% <100.00%> (ø)
torch_geometric/nn/models/dimenet_utils.py	0.00% <0.00%> (-75.52%)	⬇️
torch_geometric/nn/models/dimenet.py	14.51% <0.00%> (-53.00%)	⬇️
torch_geometric/nn/conv/utils/typing.py	81.25% <0.00%> (-17.50%)	⬇️
torch_geometric/nn/resolver.py	86.36% <0.00%> (-9.10%)	⬇️
torch_geometric/nn/inits.py	67.85% <0.00%> (-7.15%)	⬇️
torch_geometric/io/tu.py	93.90% <0.00%> (-2.44%)	⬇️
torch_geometric/nn/models/mlp.py	98.52% <0.00%> (-1.48%)	⬇️
torch_geometric/transforms/gdc.py	78.17% <0.00%> (-1.02%)	⬇️
torch_geometric/data/dataset.py	96.80% <0.00%> (-0.80%)	⬇️
... and 3 more

---

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 09f25e9...b6546fd. Read the comment docs.

[rusty1s]

Thanks @RexYing. This mostly looks good. I think edge_index_sampling should leverage SparseTensor to make this a little bit faster and to ensure that we are always operating on sorted inputs.

[rusty1s]

I think I fixed the runtime issues. One reason for this was that it is quite expensive to sample in both directions. I don't think this is really needed, and I moved to a solution where we only sample adj1 after performing the matrix multiplication. Merging this now but happy to extend/improve/make faster in later PRs.