Learning-loop-node

- fixes a bug where the creation date parameter was ignored on the socketio `detect` endpoint (49)

Detector fix:

Improve behavior when the target model is changed to speed up the change between already downloaded models. Downloaded models will only be compiled if necessary.

Detector SIO Interface:
- Add endpoints to control versioning mode and get current versioning info (`set_model_version_mode`, `get_model_version`)
- Add endpoints to get `about` info (same response as the corresponding REST endpoint)
- Add endpoints to control outbox mode (`get_outbox_mode`, `set_outbox_mode`)
- **Breaking change**: All SIO responses are now dicts. This affects the response of the `info` endpoint which now returns `{"status": "No model loaded"}` if no model is loaded.

All Nodes:
- Changes of the node status are now logged with level INFO

Further Breaking Changes:
- Enums are moved to own module

Improvements:

- Handle 429 (too many requests). When communicating with a learning loop instance via the REST interface the server may respond with 429 when it has to handle too many requests. We extended the loop communication module (used by all nodes) with a retry mechanism to retry the call several times with a delay between all attempts.

Trainer Loop Communication Changes:

- No longer send redundant fields (like the `hyperparameters`, ... ) with trainer updates (they are send with the training updates)
- Change type of `hyperparameters` to a dict.
- Add fields trainer_name and best_epoch to DC TrainingOut - used to sync the training

Trainer - Internal Changes:

- Removes the internal DC `TrainingData` and moves its fields (`image_data`, `categories`, `skipped_image_count`, `hyperparameter`) to the `Training` dataclass

Fixes:

- fixes a field name when using custom creation date