fix(gemini): handle safety-blocked metadata

[he-yufeng]

Summary

Fixes #2347.

This makes the Gemini adapter tolerate safety-blocked responses where Google returns usage_metadata with missing token counts. It also maps Gemini's SAFETY finish reason to guardrail_intervened, matching the guardrail stop reason used elsewhere in Strands.

Changes

• Default missing Gemini token counts to zero before computing usage metadata.
• Clamp computed output tokens at zero for defensive consistency.
• Map SAFETY message stops to guardrail_intervened.
• Add regression coverage for metadata-only safety-blocked responses.

To verify

Run from strands-py:

uv run pytest tests/strands/models/test_gemini.py -q --basetemp .tmp/pytest-2347 -p no:cacheprovider
uv run ruff check src/strands/models/gemini.py tests/strands/models/test_gemini.py
python -m py_compile src/strands/models/gemini.py tests/strands/models/test_gemini.py
git diff --check

Local result on Windows: 56 passed.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!