[pytorch] Add PyTorch backend integration

[fg-mindee]

Integrating PyTorch as a possible backend is a natural evolution of DocTR. Two questions need to be answered though, I added some suggestions.

How do we add the PyTorch counterpart of an object / class / function?

For module my_module.py,

• we create my_module.tf.py and put the rest in my_module.pt.py. If there are non framework-related features, they can be kept in my_module. The latter will reference the tf or pytorch implementation depending on the ibrary build
• we add a folder for each framework, and put the module named the same way in each folder for each framework.

Where do we start?

Here are my thoughts on all DocTR modules:

• datasets: the loader.py module and the call of each dataset
• documents: nothing framework-specific
• models: every module will be impacted
• transforms: the call methods will be impacted if we cannot used plain torchvision transforms
• utils: nothing framework-specific

How do we ensure that each build is running correctly?

• edit the setup.py to build the specified version
• add CI jobs to run the installation for each framework (probably will have to rethink unittests for each separate framework build)

Next steps

• Enable flexible framework selection for build & imports (feat: Added a dual framework availability checker #306, fix: Fixed framework availability check #314, chore: Ensured either TF or PT is installed and enabled #315)
• PyTorch mirror of doctr.datasets (feat: Added support of Pytorch for datasets #319)
• PyTorch mirror of doctr.models
  • doctr.models.preprocessor (feat: Added remaining components of DBNet #313)
  • doctr.models.backbones (feat: Added PyTorch support of ResNet-31 #327, feat: Added PyTorch support of VGG backbones #328)
  • doctr.models.utils (feat: Added PyTorch support to models.utils #323)
  • doctr.models.detection (feat: Added DBNet implementation in PyTorch #310, feat: Added remaining components of DBNet #313, feat: Added layer initialization in DBNet #316, feat: Added support of LinkNet in PyTorch #317)
  • doctr.models.recognition (feat: add support of CRNN in Pytorch #318, feat: added Pytorch support of MASTER #329, feat: Added PyTorch support of SAR #333, fix: bug in MASTER loss #335, feat: add loss + MASTERPostProcessor in Pytorch #340, refactor: Refactored MASTER logits decoding #342)
• PyTorch mirror of doctr.transforms (feat: Added remaining components of DBNet #313, feat: Added support of remaining transforms in PyTorch #322)

[fg-mindee]

Closing this since ONNX support and reference script are a bit far off from this PR