PoC: assisted inpaint

[jokester]

https://github.com/zyddnys/manga-image-translator has decent CNN models we can steal from

[jokester]

see also: https://frederik-uni.github.io/manga-image-translator-rust/modules/inpainters.html

[jokester]

also: multimodal LLM (with less dof of fonts / typesetting)

[jokester]

Qwen-Image-Edit and nano banana worked well on inpainting (removal of texts and balloons) task.

btw asking them to translate whole page e2e doesn't work.