Enable index_sort

[DamianSzwichtenberg]

Enable index_sort introduced in pyg-lib.

Till now, sorting indices (1d tensor) have been working sequentially. This PR enables the index_sort operation, which makes such a sort work in parallel (parallelized radix sort).
Related PR: pytorch_sparse:306

In training benchmarks, data set load time was improved on average by 3.82 times, and training time by 1.06 times.

[codecov]

Codecov Report

Merging #6554 (83f9c3d) into master (57d107f) will increase coverage by 0.29%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #6554      +/-   ##
==========================================
+ Coverage   85.92%   86.21%   +0.29%     
==========================================
  Files         412      413       +1     
  Lines       22784    22750      -34     
==========================================
+ Hits        19577    19615      +38     
+ Misses       3207     3135      -72     

Impacted Files	Coverage Δ
torch_geometric/profile/profile.py	93.65% <ø> (+60.31%)	⬆️
torch_geometric/data/graph_store.py	90.40% <100.00%> (+0.07%)	⬆️
torch_geometric/nn/conv/rgcn_conv.py	95.36% <100.00%> (ø)
torch_geometric/nn/dense/linear.py	89.13% <100.00%> (+0.07%)	⬆️
torch_geometric/sampler/utils.py	92.00% <100.00%> (+0.16%)	⬆️
torch_geometric/utils/__init__.py	100.00% <100.00%> (ø)
torch_geometric/utils/coalesce.py	91.30% <100.00%> (ø)
torch_geometric/utils/sort.py	100.00% <100.00%> (ø)
torch_geometric/utils/sort_edge_index.py	88.00% <100.00%> (+0.50%)	⬆️
torch_geometric/nn/conv/appnp.py	100.00% <0.00%> (+6.55%)	⬆️
... and 1 more

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

[LukeLIN-web]

Why I met problem #6563