Support for reprocessing detections and skipping detector

[mihow]

The slowest part of the processing pipeline is running the object detector on the raw source images from cameras. This is required once, but should be skipped whenever possible. For example, when running a new classifier on existing detections or adding new data like feature vectors (for #751).

Implementation:

• Update the Antenna API for the ML backend to send existing detections.
• Update the ADC to determine if detection or classification is needed.

This is also important because

1. Some detections & occurrences are missing classifications now. This should fill those in without re-running source images.
2. We are saving all softmax scores in the DB, but existing detections don't have them. This should fill in those.
3. When we add feature vectors to the responses, this can fill in those fields.

Some pipelines could also be configured to return results from multiple classifiers in one call. We already return the binary & fine-grained results in one call since #684. For example the global classifier predictions could be returned with regional model results.

[vanessavmac]

DONE

Immediate TODO: in ml backend example, fix the schemas; main idea is request and response schemas must be identical between antenna and the processing services (add back the crop image url to avoid issues in prod)

• GOAL: user processes empty source image with a pipeline; this will produce detections (potentially with classifications); any time after, using any pipeline, should NEVER run the detector again ONLY run the classifier (always)

Example of multiple classifications: https://antenna.insectai.org/projects/18/occurrences/559687?algorithm=11&tab=identification --> this is a good test case (the detections have multiple classifications; only ever one detector in prod)

• Something to try: use zero shot then the random classifier to get actual butterfly species