Fix output metadata processing in loaders 'filter_fn()` for multinode

[JakubPietrakIntel]

The output of DistributedNeighborSampler._collate_fn() includes nfeats, nlabels, efeats which are our metadata[-3], metadata[-2] and metadata[-1]. Here: L717C1-L718C22

[codecov]

Codecov Report

Attention: 8 lines in your changes are missing coverage. Please review.

Comparison is base (ccbbbdd) 88.42% compared to head (f32b554) 88.39%.
Report is 2 commits behind head on master.

❗ Current head f32b554 differs from pull request most recent head d1b901b. Consider uploading reports for the commit d1b901b to get more accurate results

Files	Patch %	Lines
torch_geometric/loader/link_loader.py	0.00%	5 Missing ⚠️
torch_geometric/loader/node_loader.py	40.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #8377      +/-   ##
==========================================
- Coverage   88.42%   88.39%   -0.03%     
==========================================
  Files         475      475              
  Lines       28839    28843       +4     
==========================================
- Hits        25500    25496       -4     
- Misses       3339     3347       +8     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[rusty1s]

Thanks. I misunderstood the implementation, sorry :)

From a PyG interface perspective, I personally think it is quite confusing that SamplerOutput also returns the features, but I see that it makes sense to fetch the features during sampling. However, this does not really seem to generalize to more feature attributes (e.g., time), so let me know if there exists plans to extend this to more general cases.

[JakubPietrakIntel]

No worries :) thanks for correcting that quickly. You're right, we will need revise our Sampler output to be more customizable. Remark noted!