Learning-loop-node

Detector enhancements:
- enhance feedback in upload EP
- make confidence thresholds configurable

Fixes bug in upload EP

New features:
- SIO endpoint 'restart' that triggers a container-restart
- Trainer now deletes old trainings on drive (for the same project) on start of a new training

Improvements:
- Detector endpoint `upload` now also handles detections and can receive custom tags
- Trainer triggers sio reconnection on timeouts

Fixes:
- Detector can now do detections even if learning loop is not reachable

- fixes a problem with uploading point detections

Fix a bug where detectors without credentials endlessly try to connect to the server.

This release largely improves the stability of the communication with the learning loop backend. BaseModels were changed to dataclasses to speed up the generation of objects. The codebase is now mostly type-annotated and uses common formating schemes.