Feature/multi bbox predictions

[mualicemu]

Description

This PR introduces a new execution pathway (multi_predict) to extract and return independent species classifications for every bounding box detected in an image, bypassing the legacy constraint that flattened all detections into a single prediction.

Key Changes

1. classifier.py: Added multi_preprocess() to dynamically generate separate image crops for each detected bounding box within the image.
2. ensemble_prediction_combiner.py: Added combine_predictions_for_all_items(), a new 1-to-many combiner loop. Instead of hardcoding detections[0], it iterates over all bounding boxes, applies geofencing/roll-up rules individually, and retains the physical bounding box [ymin, xmin, ymax, xmax] coordinates in the output dictionary.
3. ensemble.py: Added multi_combine() to orchestrate the new zippering logic.
4. multiprocessing.py: Appended a new multi_predict() method to the SpeciesNet class. This safely batches the cropped tensors on the GPU and invokes the new combiner.

Note: Strict backwards compatibility is maintained. The original predict() method and combine_predictions_for_single_item remain completely untouched for legacy integrations.