Set of commits to be tagged v0.20

Changelog

Features

The main new feature is a REST API, so that LarvaTagger can be run as either a frontend or a backend. The backend contains and operates the tagging backends, while the frontend serves the Web UI.

• REST, REST.Server, REST.Client and REST.Model modules
• backend switch for the scripts/larvatagger.sh script
• Tagging backends and taggers (“model instances”) can contain a metadata or metadata.json JSON file along with their files (first step towards #59)
• Tagging backends can specify whether switches among train, finetune, predict and embed require the make_dataset step or not (Taggers.apply_make_dataset)
• [WebUI] Contextual help for tagging backends and taggers.

Bug fixes

• [WebUI] The Autotag button is disabled if no data files are loaded
• scripts/larvatagger.sh build --dev and the recipes/Dockerfile.local file are usable again

Other changes

• LarvaTagger.jl no longer passes --skip-make-dataset by default to the tagging-backend Python program for the predict and embed switches (this was already the default in tagging-backend)
• Calls to tagging-backend are logged at INFO level
• [WebUI] NyxWidgets upgraded to 0.2.2, and upload file/delete file/create directory buttons enabled by default with --server-url (implements part of #8)
• The scripts/install.sh script selects the main branch of LarvaTagger.jl by default; --with-default-backend now selects the Python3.11/Torch2 variant; the former Python3.8/Torch1 variant can be selected with --with-backend --legacy