dataclasses and code speedup

[ali-izhar]

simplified code structure api and experiment module adding dataclasses, dynamic model registry for torchvision models and ultralytics, dynamic inference utils.

[nbovee]

Pretty happy with what we have so far, current notes:

• local only fallback on device_config.yaml needed so we can run simple integration tests before being fully configured
• similar fallbacks for defaults on other scripts within ./tests
• high change I introduced a lot of off-by-one errors in the hook changes since I couldn't test them without full configs today
• looking into still keeping this installable, but adding a simple cli command to dump configs in current directory so we can still use pypi
• if python 3.11 can be we used as our lower bound we could swap configs to toml files with stdlib additions
• uv.lock tends to need de-duping and locks up, for the time being I have been doing uv lock | uv sync --extra cpu to clean and rebuild it

[nbovee]

just read further into our template files, now I know why you though I was crazy asking for the current configs!

[nbovee]

Merging in to staging, rolling into main after checking readmes