RAFT model and training reference

[NicolasHug]

I'm opening this issue to keep track of the progress made on the RAFT implementation. Implementing RAFT is part of our H2 goals: #4187

• model architecture (normal and "small" one) Add RAFT model for optical flow #5022

• new datasets

  • Flying Chairs Add FlyingChairs dataset for optical flow #4860
  • FlyingThings3D Add FlyingThings3D dataset for optical flow #4858
  • Sintel Add Kitti and Sintel datasets for optical flow #4845
  • KittiFlow Add Kitti and Sintel datasets for optical flow #4845
    • support for 16bits pngs for KittiFlow Add support for 16 bits png images #4657
  • HD1K (maybe) Add HD1K dataset for optical flow #4890

• new transforms and augmentations Add transforms and presets for optical flow models #5026

• training and evaluation reference for optical flow models Add training reference for optical flow models #5027

• pretrained weights (our own weights from training recipe + porting of original weights from paper):

  • C+T: Add pretrained weights on Chairs and Things for raft_large #5060
  • C+T (small) Add pretrained weights for raft_small from original paper #5070 Add our weights for raft_small on C+T #5076
  • C+T + Sintel fine-tuning Pre-trained weights for raft-large, fine-tuned on sintel #5079
  • C+T + Sintel fine-tuning + Kitti fine-tuning Add raft_large weights fined-tuned on Kitti #5081

• Bonus: implement new prototype API Add raft builders and presets in prototypes #5043

The support for Kitti would require support for 16bits png, which we don't have at the moment. EDIT: we do now

cc @pmeier @datumbox