[models] Add detection & recognition models with much lighter backbones

[fg-mindee]

Comparing similar vision tasks, the segmentation part of text detection is easier than real-life scene segmentation (on Cityscapes for instance) which uses ResNet-like backbones, and even easier than medical segmentation which uses often lighter feature extractor (cf. UNet).

Additionally, the feature extraction of text recognition is similar to image classification for a single character image (MNIST but with more classes) which uses very light backbones.

So here is a suggestion:

• train a very light image classifier for character classification on a specific vocab (it will be used as a pretrained backbone for text recognition) (feat: Added pretrained PyTorch mobilenets #415 feat: add mobilenet weights for character classification (TF) #421)
• train a text recognition model with this (feat: Added rectangular stride MobileNets #483, feat: add pytorch ckpts for crnn & mobilenet_v3_large #487, feat: add crnn_mobilenet_v3_small tf ckpts #517, feat: add crnn_mobilenet small & large ckpts for torch backend #516)
• train a light text segmentation model (feat: add db_mobilenet_v3_large (TF) ckpt & benchmark #485, feat: add pytorch ckpts for crnn & mobilenet_v3_large #487)

This means that DocTR will extend the list of supported tasks by adding:

• character classification (feat: Added character classification training scripts #414)
• text segmentation

[fg-mindee]

Closed by #516 & #517