[DataLoader] Typing Enforcement for DataPipe at Compile-time (#54020)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54020
## Feature
- Add `issubtype` to check the type is a subtype of the other type.
- Add `_DataPipeMeta` (mimic Python typing 3.6)
- Add `type` attribute for each DataPipe
- Save original `__init__` function for each DataPipe
- Validate return hint of `__iter__`
- Replace `__init__` function bases on `type`
- Fixed type: Put original `__init__` back, if it exists or use a plain `__init__`
- Non-fixed type: Add new `__init__` with the functionality to copy `cls.type` for each instance. (Optimized for memory)
No Error for main repo, `torchvision`, `torchaudio` and `torchtext`.
## Future
- Add same thing for `__getitem__`.
- When DataFrame came out, add an another type for DataFrame with column name and type.
Test Plan: Imported from OSS
Reviewed By: VitalyFedyunin
Differential Revision: D27327232
Pulled By: ejguan
fbshipit-source-id: fd3a6029c16f5d814b1d7e1b1566fdcd8fd1ad9a
