Parse Gemini 2.5 native object-detection format in vlm_as_detector

[dkosowski87]

What does this PR do?

After migrating the default Gemini model from gemini-2.0-flash to gemini-2.5-flash (#2395), hosted E2E tests showed that Gemini object-detection workflows still completed but returned parsed_prediction: null. Gemini 2.5 often responds with its native bounding-box JSON (box_2d + label, 0–1000 scale, top-level array) instead of the legacy format the vlm_as_detector parser expected ({"detections": [{"x_min", "class_name", ...}]} with 0–1 coordinates).

This PR adds shared Gemini object-detection parsing that auto-detects both output shapes and converts them into standard sv.Detections. The parser accepts a JSON root that is either a dict with a detections key or a top-level list, maps label / class / class_name to class IDs, and converts box_2d coordinates to pixel-space xyxy boxes. Both vlm_as_detector@v1 and @v2 route google-gemini object-detection output through this shared logic; OpenAI and Claude continue to use the existing legacy parser unchanged.

Main elements:

• New gemini_detection_parsing.py shared module for Gemini detection response normalization
• Updated vlm_as_detector@v1 and @v2 JSON parsing to accept list or dict roots
• Unit tests for Gemini 2.5 native box_2d output in test_v1.py and test_v2.py

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Other:

Testing

Unit tests

• Legacy Gemini/Claude detection JSON (x_min / class_name format) — existing coverage unchanged
• Gemini 2.5 native detection JSON (box_2d / label, top-level array) — new tests in test_v1.py and test_v2.py
• Invalid/malformed JSON still returns error_status: True and predictions: None

Integration tests

• None for this PR. Fixes the hosted-platform scenario exercised by tests/inference/hosted_platform_tests/workflows_examples/test_workflow_with_gemini.py::test_object_detection_workflow once deployed.

Other

.venv/bin/python -m pytest tests/workflows/unit_tests/core_steps/formatters/vlm_as_detector/ -q
# 30 passed

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context

• Symptom before fix: test_object_detection_workflow failed with TypeError: 'NoneType' object is not subscriptable because vlm_as_detector returned predictions: None while the Gemini step succeeded.
• Related change: Complements refactor(gemini): remove deprecated model versions and update default to 2.5-flash #2395 (Gemini 2.0 deprecation / default bump to 2.5-flash). No changes to google_gemini blocks — they still return raw model text; parsing stays in the formatter.
• Follow-up (optional): Consider adding response_schema to google_gemini object-detection prompts to enforce a single JSON shape at the API level, reducing reliance on format auto-detection.