Add Gemini 3.6 and 3.7 structured VLM output support

[Xayar145]

Before submitting

• Self-reviewed the code
• Updated documentation in Google style
• Added/updated tests
• All tests pass locally

Description

Adds Google Gemini 3.6 and 3.7 adapters to Detections.from_vlm for the current structured object-detection and segmentation response format.

Type of Change

• ✨ New feature (non-breaking change which adds functionality)

Motivation and Context

The existing Gemini adapters stop at 3.5 and expect the legacy top-level array and encoded-mask format. Gemini 3.6 and 3.7 structured responses wrap items in a top-level boxes object and represent segmentation masks as normalized polygons. Without dedicated VLM members and parsing, Detections.from_vlm rejects these model variants.

Format reference: https://ai.google.dev/gemini-api/docs/image-understanding#object-detection

Changes Made

• Register GOOGLE_GEMINI_3_6 and GOOGLE_GEMINI_3_7 with result and argument validation.
• Parse structured boxes, optional confidence values, class filters, and normalized polygon masks into Detections.
• Preserve valid empty-mask shapes when class filtering removes every item, and document the new adapters in the capability tables and changelog.

Testing

• I have tested this code locally

• I have added unit tests that prove the feature works

• All new and existing tests pass

• Focused VLM suite: 114 passed

• Full suite: 3617 passed, 20 skipped

• Pre-commit: all hooks passed, including Ruff, mypy, codespell, and Markdown formatting

Additional Notes

The 3.7 adapter reuses the 3.6 structured parser because both versions expose the same box and polygon schema.

[CLAassistant]

All committers have signed the CLA.