Update GraphStore and FeatureStore [1/6]

[kgajdamo]

This code belongs to the part of the whole distributed training for PyG.

Please be aware that this PR should be merged before Loaders package! - @JakubPietrakIntel
Loaders:

1. Add base class DistLoader #8079
2. Add DistributedNeighborLoader [3/6] #8080
3. Add DistributedLinkNeighborLoader [4/6] #8085

Other PRs related to this module:
DistSampler: #7974

[codecov]

Codecov Report

Merging #8083 (70a5594) into master (3eafd5d) will decrease coverage by 0.30%.
The diff coverage is 60.09%.

@@            Coverage Diff             @@
##           master    #8083      +/-   ##
==========================================
- Coverage   87.41%   87.12%   -0.30%     
==========================================
  Files         473      473              
  Lines       28648    28724      +76     
==========================================
- Hits        25043    25025      -18     
- Misses       3605     3699      +94     

Files	Coverage Δ
...rch_geometric/distributed/dist_neighbor_sampler.py	0.00% <0.00%> (ø)
torch_geometric/distributed/local_graph_store.py	95.87% <93.75%> (+4.66%)	⬆️
torch_geometric/data/graph_store.py	92.75% <76.31%> (-5.61%)	⬇️
torch_geometric/distributed/utils.py	51.94% <23.52%> (-7.43%)	⬇️
torch_geometric/distributed/partition.py	83.22% <56.66%> (-15.95%)	⬇️
torch_geometric/distributed/local_feature_store.py	57.85% <45.45%> (-22.41%)	⬇️

... and 1 file with indirect coverage changes

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

[JakubPietrakIntel]

@rusty1s @ZhengHongming888
I added ac49add which enforces sorting by destination node in LGS edge_index (in other words sort by col, or edge_index[1]) to match CSC format used for sampling. This is condition ensures that input data won't be permuted again at neighbor sampler init.
We do that to make sure that correct edge_ids are returned at sampler's output and in filter_fn in NodeLoader i.e. L161 the expression is evaluated correctly with perm = None.
We have considered alternative solutions and decided that this one will be the best regarding performance, at a cost of adding a constraint on incoming data. We can't have a global perm stored in LGS, as METIS is not performing a simple split, so the ordering would be lost. The other solution would be to use local values per partition in edge_id, hence increasing complexity and adding yet another re-indexing step. Furthermore, it would require RPC requests for perm values located in remote partitions, because the sampler may return edge_ids located in remote partition.