fix: fingerprint cacheable context prefix

[he-yufeng]

Summary

• Use the cacheable content prefix for fingerprint-only cache metadata instead of the full request contents.
• Keep request-scoped dynamic instruction contents out of the initial fingerprint chain.
• Update CacheMetadata.contents_count docs to match the prefix semantics.

Fixes #6062.

Testing plan

This PR is covered by focused unit tests for the Gemini context cache manager. The important behavior is that cache fingerprints are derived only from the cacheable prefix, while request-local dynamic instructions still flow into the request without changing the cache identity.

Verification output

Run locally on Windows with Python 3.13:

python -m py_compile src\google\adk\models\gemini_context_cache_manager.py src\google\adk\models\cache_metadata.py
# passed

PYTHONPATH=src python -m pytest tests\unittests\agents\test_gemini_context_cache_manager.py -q
# passed

python -m ruff check src\google\adk\models\gemini_context_cache_manager.py src\google\adk\models\cache_metadata.py tests\unittests\agents\test_gemini_context_cache_manager.py
# passed

git diff --check
# passed

Risk

Low. The patch narrows what contributes to fingerprint-only cache metadata; it does not change response generation, model selection, or live cache service calls.

[adk-bot]

Response from ADK Triaging Agent

Hello @he-yufeng, thank you for submitting this pull request!

This is a bug fix (associated with #6062). To help reviewers process your contribution more efficiently, could you please update the PR to include:

1. A testing plan section describing how you tested your changes (as requested in our Contribution Guidelines).
2. Logs or verification output (such as the output of the passed pytest results) showing that the tests run successfully after your changes.

Thank you for your help in keeping the codebase robust!