Properly support dynamic inheritance in Batch

[mananshah99]

This diff properly supports dynamic inheritance with the DynamicInheritance metaclass in Batch. The initial issue was that of conflicting metaclasses: abc.ABCMeta and DynamicInheritance for Data and HeteroData objects. The resolution required two steps:

• Metaclass conflict resolution of the aforementioned metaclasses
• Patching pin_memory in PyTorch which makes assumptions about MutableMapping before checking whether pin_memory exists in the class. This in turn had adverse effects of calling the metaclass of the class again when constructing the new class as part of pin_memory, which resulted in the DataDataBatch problem observed in the filed issue.

Hopefully the latter fix can be deprecated in the future, with a patch to PyTorch (see pytorch/pytorch#82185)

Closes #4848

[codecov]

Codecov Report

Merging #5051 (5b4027e) into master (0e2d987) will decrease coverage by 0.01%.
The diff coverage is 50.00%.

❗ Current head 5b4027e differs from pull request most recent head c4507eb. Consider uploading reports for the commit c4507eb to get more accurate results

@@            Coverage Diff             @@
##           master    #5051      +/-   ##
==========================================
- Coverage   82.90%   82.89%   -0.02%     
==========================================
  Files         331      331              
  Lines       18150    18160      +10     
==========================================
+ Hits        15048    15053       +5     
- Misses       3102     3107       +5     

Impacted Files	Coverage Δ
torch_geometric/loader/dataloader.py	74.46% <50.00%> (-6.62%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us.

[rusty1s]

Amazing, didn’t know it is due to pin_memory. Can we add a test which might have failed without this patch?

[mananshah99]

Can we add a test which might have failed without this patch?

Since this is an issue with pin_memory, it will be hard to verify this on CPU since torch _BaseDataLoaderIter internally sets pin_memory to False without CUDA. We could define a custom dataloader iterator that overrides this functionality and write a test for this, but that might be overkill :)