fix(leaderboard): align GPT-5.5 Long results and evaluation notes

[ye1457]

Update the leaderboard assets to match the published evaluation results and make their qualifications easy to find.

• Correct the GPT-5.5 LIBERO Long results to 70/100 and 55/100, with the eight-suite average rounded to 75.13%.
• Keep the four independent starred evaluation notes, including Flash memory/perception/seed selection and the two frozen Astra memory batches.
• Link the relevant model names to these notes in charts, full tables and the standalone fallback. The links support keyboard navigation and expose accessible labels.
• Style scope-note in the documentation fallback and remove the unused benchmark-method selector.

The latest note-link changes leave results.json, overview images and CSV contents unchanged. Validation covered both languages, desktop/mobile, light/dark, offline and no-JavaScript pages; all 28 compared CSV outputs matched byte-for-byte.

Companion documentation: RLinf/RPent#215. Once this PR is merged, that PR will pin the final main commit and verify the public assets before merging.

[ye1457]

Companion documentation PR: RLinf/RPent#215 . Its resource references are pinned to 705bd44. Public jsDelivr byte checks and unmodified Sphinx integration passed all 128 checks across both languages, desktop/mobile, light/dark, CSV, native fallback and separate cost-page routing. Merge this resource PR first; no automatic merge.