[models] TF & PT: Retrain all detection models with random rotation augmentation

[felixdittrich92]

🚀 The feature

All detection models should be trained with rotation augmentation

PyTorch:

• db_resnet50
• db_resnet34
• db_mobilenet_v3_large
• linknet_resnet18
• linknet_resnet34
• linknet_resnet50

TensorFlow:
Note: (#1348 should be merged before we start with TF)

• db_resnet50
• db_mobilenet_v3_large (running)
• linknet_resnet18
• linknet_resnet34
• linknet_resnet50

Motivation, pitch

• After this retrain every detection model can be used also for rotated pages / text detection
• We can drop linknet_resnet18_rotation (TF) and db_resnet50_rotation (PT)
• We can remove the limitation ->

  doctr/doctr/models/detection/zoo.py

  Line 40 in aadb96a

  if arch not in ROT_ARCHS and not assume_straight_pages:

• with the new trained models and [prototype] compute orientation on segmentation map #1336 the page orientation detection will be much more robust (range -90 - 90 degrees) / upside down still open