Add AIME26 benchmark

[taivu1998]

Summary

• Add an AIME26 custom benchmark adapter following the existing AIME25 prompt, repeat, boxed-answer extraction, and is_equiv scoring pattern.
• Vendor eval/chat_benchmarks/AIME26/data/aime26.json with 30 AIME 2026 problems from MathArena, mapped as I-1 through I-15 and II-1 through II-15.
• Document AIME26 in the README benchmark list and add a focused stdlib-only data integrity test.

Data Source and License

• Source dataset: https://huggingface.co/datasets/MathArena/aime_2026
• License reported by the source dataset: CC BY-NC-SA 4.0.
• The vendored JSONL keeps the same lightweight local schema used by AIME25: id, problem, solution, answer, and url.
• solution is intentionally left blank because this PR only vendors problem statements and final answers.

Validation

• uvx --with pytest pytest tests/test_aime26_data.py passed: 3 tests.
• python3 -m py_compile eval/chat_benchmarks/AIME26/eval_instruct.py tests/test_aime26_data.py passed.
• Local AIME26 data invariants passed: 30 rows, ordered unique IDs, expected answer sequence, numeric answers in the AIME 0-999 domain, blank solutions, source URL present on every row.
• Live source parity was checked against freshly downloaded MathArena/aime_2026 rows: local problems and answers match exactly.
• The first 15 rows were checked against MathArena/aime_2026_I and match exactly.
• A stubbed TaskManager(task_list=["AIME26"]) discovery and scoring smoke test passed.
• git diff --check passed.

Notes

• Full uv run python ... import validation is blocked before reaching AIME26 by the existing project metadata issue at pyproject.toml: fschat @ file:eval/chat_benchmarks/MTBench.
• I did not find a public official MAA problem/answer source during the source research pass, so this PR should be treated as MathArena-sourced rather than MAA-officially verified.