Remake localtagger with imgutils for more models

[AshtakaOOf]

Localtagger is the backend of Farterrogator which is a material theme frontend for tagging images with danbooru tags.

Goal: More easily maintainable backend for adding new models etc with dghs-imgutils and the existing Kaloscope code for artist similarity.

dghs-imgutils is available as a pypi wheel https://pypi.org/project/dghs-imgutils/
this will be used to inference all the models we'll be adding as part of the remake (it can do dbv4 wd camie)

Models:

• Animetimm dbv4, new family of model similar to the seasoned wd taggers
  • https://huggingface.co/animetimm/mobilenetv4_conv_aa_large.dbv4-full, instant tagger
  • https://huggingface.co/animetimm/swinv2_base_window8_256.dbv4-full, fast and good
  • https://huggingface.co/animetimm/caformer_b36.dbv4-full, balanced and recommended by animetimm
  • https://huggingface.co/animetimm/eva02_large_patch14_448.dbv4-full, pretty big but amazing
  • https://huggingface.co/animetimm/convnextv2_huge.dbv4-full, huge but deadly accurate
• WD Tagger v3, seasoned models from 2024
  • https://huggingface.co/SmilingWolf/wd-swinv2-tagger-v3, fast and balanced wd
  • https://huggingface.co/SmilingWolf/wd-eva02-large-tagger-v3, best wd tagger
• Camie v2, different feel and more inconsistent
  • https://huggingface.co/Camais03/camie-tagger-v2

note1: animetimm's repos are locked behind agreeing to share email
note2: we cannot use onnx for animetimm as they are only packaged as safetensors as such use pytorch
note3: camie can guess artist but it's bad at it so omit the feature

Preferably the code should be well structured and not plopped into one file.
Also if possible easily configurable to choose where the models are hosted etc.

Goodies (optional): Send model list to frontend so backend doesn't have to updated to use newly added models. Ask developer how to tackle i18n if this is done.

Post remake: update frontend to the new api so the backend is actually usable from it.