[PierreRaybaut]

Closes #89

Summary

Sigima 1.4 selects the pixels of image regions of interest by their centers (#83, #84). This changes the result of almost every image operation when its input carries regions of interest, while calls on images without ROIs are unchanged. Increasing every image contract (~117) would refuse the replay of all image activities recorded with 1.3; keeping them unchanged let such replays report different without telling why.

The convention is now a rule of the execution context, like the X-alignment policy:

• sigima.image.roi.pixel_centers, version 1 (sigima.proc.contracts.IMAGE_ROI_CONVENTION, read-only mapping);
• image_roi_convention(inputs): the record when an input image carries (non-empty) regions of interest, None otherwise;
• describe_call returns a new context mapping ({"roi_convention": {...}} when it applies), and prepare_execution adds it to the prepared context;
• for a replay, prepare_execution(..., context=...) raises ExecutionError("unsupported_context") when an input image carries regions of interest and the recorded rule is missing (recorded before Sigima 1.4: explicit message) or is another one.

Calls on images without ROIs and on signals are not concerned. DataLab-Capsule records the context mapping of the description since DataLab-Platform/DataLab-Capsule#4 (merged).

Docs: module docstrings, guidelines (rules shared by all operations are versioned in the context, not in the contracts), release note (with French translations).

Tests

• New test_image_roi_convention_is_recorded_and_checked: nothing recorded nor required without ROI (or with an empty ROI), recorded by the description and the preparation with a ROI (also on an operand), replay with the record gives the same result, replays without it or with another rule/version are refused, signal ROIs not concerned.
• Full suite: 2313 passed; --qualification-report: no problem.
• DataLab (develop with DataLab-Capsule develop): provenance, capsule and replay tests pass (150). End-to-end check: an image with a ROI processed (Gaussian filter, centroid) records the rule and replays exactly; the same activity without the rule is refused with the explicit reason; an image without ROI records nothing.