Validate parent_id alignment between images and predictions in Dynamic Crop

[rs-03]

What changed

DynamicCropBlockV1.run pairs images with predictions positionally via zip(images, predictions). When those two inputs are produced by independent workflow branches they can be misaligned, so a crop gets taken from one image using another image's detections, producing silently incorrect results (the "nonsense results" described in the issue).

This adds a contract check in crop_image: when the predictions carry a parent_id (as model-produced sv.Detections do), it must match the image's parent_metadata.parent_id. On a mismatch the block now raises a clear ValueError explaining the misalignment instead of returning wrong crops. The check mirrors the block's existing detection_id contract validation and is skipped when predictions carry no parent_id, so behavior is unchanged for correctly aligned inputs.

Closes #390

Design note

I implemented this as validation with an explicit error, consistent with the block's existing detection_id contract check, because a silent realignment would hide a genuinely misconfigured workflow rather than surface it. If you would instead prefer the block to match predictions to images by parent_id, I am happy to switch to that approach.

Tests

Added two unit tests in tests/workflows/unit_tests/core_steps/transformations/test_crop.py:

• test_crop_image_raises_when_predictions_parent_id_does_not_match_image (fails on the pre-fix code, confirming the guard is effective)
• test_crop_image_accepts_predictions_whose_parent_id_matches_image

Command run:

pytest tests/workflows/unit_tests/core_steps/transformations/test_crop.py

Result: 47 passed. Files were formatted with black (26.3.1) and isort (5.13.2); both leave the changes unchanged.