Quantize CV models

[mkopcins]

Quantize CV models to improve performance, especially for real time usecases

Image Embeddings:

• clip-vit-base-patch32-vision

OCR

• Craft Detector
• Craft Recognizer (Allegedly, quantized version was much worse in quality with no performance benefits.)

Object Detection

• ssdlite320-mobilenet-v3-large

Style Transfer

• Candy
• Mosaic
• Rain Princess
• Udnie

Classification

• efficientnet-v2-s

Semantic Segmentation

@barhanc could you export quantized versions together with the non-quantized ones?

• DeeplabV3
• LRASPP
• FCN

Text to Image

Our current diffusion model is quite bad and it's not worth to quantize it right now. In the future we can export better models e.g. https://huggingface.co/IDKiro/sdxs-512-dreamshaper and then return to quantizing it. #939

• Text Encoder (?)
• UNET
• VAE

[msluszniak]

As a follow up to this one, we need to add versioning to our export flows that will mimic versions of this library.

[msluszniak]

@barhanc Regarding Text to Image, could you make a separate isseu for it? You can just copy the text from here.