[OlympiadBench] Add separate official English text-only task

[taivu1998]

Add OlympiadBench_official alongside the existing OlympiadBench task, whose implementation and scoring behavior remain unchanged. This is deliberately a separate benchmark identity rather than a bug-fix rewrite: changing the dataset population, prompt, and grader in place would silently make new OlympiadBench scores incomparable with existing results.

Compared with OlympiadBench on main, which defaults to a bundled 30-problem sample, a generic boxed-answer prompt, and correct-if-any hendrycks_math grading (current adapter, current grader), OlympiadBench_official:

• evaluates exactly 910 English open-ended text-only problems: 674 mathematics rows from OE_TO_maths_en_COMP and 236 physics rows from OE_TO_physics_en_COMP. The authors link this updated dataset release and define OE, TO, and en as open-ended, text-only, and English (dataset release, subset taxonomy);
• adapts the authors' English open-ended prompt, including answer-type wording, ordered multiple answers, and unit placement (prompt implementation);
• vendors the authors' automated scorer under its MIT license, preserving complete multi-answer grading and per-problem tolerances for numerical, expression, equation, interval, percentage, and plus/minus equivalence (reference scorer); and
• pins both dataset and reference-code revisions, rejects scope/schema/row-count drift, records protocol metadata, reports aggregate and Math/Physics metrics, and supports native pass@k.

The separate task name preserves historical compatibility while giving users an explicit, reproducible path for the authors' English text-only protocol. It includes its own single-task config and install extra, and both OlympiadBench variants remain importable in the torch-free lean installation.