Add Exo7 French MCQ math benchmark

[Lduignan1]

Summary

Adds community_tasks/exo7.py: Exo7, a multi-label multiple-choice math benchmark for French undergraduate students (dataset: OpenLLM-BPI/Exo7MCQ, source: http://exo7.emath.fr/). Many items have more than one correct answer.

Two zero-shot scoring paths are exposed:

• Logprob path (MCF and Hybrid formulations): TruthfulQA MC2-style probability-mass metric (Exo7MCMetric) over length-normalized choice logprobs (CF not included as many questions do not include context required to answer e.g. "Quels sont les propositions vraies ?")..
• Generative path: model emits Réponse : A, C; scored with set-F1 and exact-set-match. Includes a \boxed{...} answer-format fallback.

Test plan

• Task is discoverable via the lighteval task registry and imports cleanly.
• Run the Exo7 task end-to-end on a small instruct model in both logprob (MCF / Hybrid) and generative variants; confirm metrics are produced.