Telescopetest-io: performance and "accuracy" improvements to AI content-rating

To improve performance of AI content-rating, we should use a KV binding to store R2 asset ratings. This decision was made based off the discussion in a previous PR: https://github.com/cloudflare/telescope/pull/207#discussion_r2966917173.

To improve rating "accuracy", should also see if we can send data in `metrics.json` or the test title and description to the test-rating as well.