Add filesystem abstraction using fsspec

[asherbondy]

Changes made:

• Add fs_utils.py for API that offers 1:1 replacement for current filesystem usage.
• Update usages of open, glob, makedirs, normpath to use new API.
• Update test/datasets/* to use in-memory filesystem, simplifying test and testing fsspec API.

[asherbondy]

Hey folks, Just in case it is not clear from the PR title. I would like to make sure the proposed API is agreed upon based on the presented cases before attempting to carry the fs abstraction throughout the rest of the code base.

[codecov]

Codecov Report

Attention: 10 lines in your changes are missing coverage. Please review.

Comparison is base (4f5100f) 88.32% compared to head (a023f20) 87.66%.

❗ Current head a023f20 differs from pull request most recent head a42cf72. Consider uploading reports for the commit a42cf72 to get more accurate results

Files	Patch %	Lines
torch_geometric/io/fs.py	86.48%	10 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #8379      +/-   ##
==========================================
- Coverage   88.32%   87.66%   -0.66%     
==========================================
  Files         475      476       +1     
  Lines       28942    29011      +69     
==========================================
- Hits        25562    25433     -129     
- Misses       3380     3578     +198     

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

[rusty1s]

Thanks. I think this is a great start. I moved the files to torch_geometric.io and suggest usage of

from torch_geometric.io import fs

I think this is a bit cleaner.

The cp function for my taste seems a bit overloaded. Is it common practice to fuse downloading and extraction here? I think this is a bit error-prone. Made some updates but broke memory:// support in conftest, since fs.ls doesn't return the protocol :(