Add AIME26 task

[Vedant-Agarwal]

Closes #1167

Adds aime26, aime26_avg, and aime26_gpassk task configs to src/lighteval/tasks/tasks/aime.py, mirroring the aime25 trio exactly (same prompt function, record_to_sample, and metric choices) — only the dataset repo, splits, and version differ.

Dataset: math-ai/aime26 (confirmed in the issue comments), config default. One adaptation from the aime25 pattern: this dataset ships a single test split (30 problems) rather than train, so hf_avail_splits/evaluation_splits are ["test"]. Columns (problem, answer) match what aime_prompt expects — verified by loading the live dataset and running the prompt function on real rows.

Unlike the earlier attempt in #1218 (which targeted the multilingual LumiOpen/mAIME* datasets and bundled unrelated CI/dependency changes), this is a minimal single-file change doing exactly what the issue asks.

Verification:

• Registry resolves all three tasks with correct metrics
• pytest tests/unit/tasks/test_registry.py → 8 passed
• ruff check / ruff format --check clean

[Vedant-Agarwal]

Friendly ping @NathanHB — this adds the AIME26 task (closes #1167), mirroring the existing aime25 configs exactly, so it's a small isolated addition. It's mergeable and I am happy to adjust if you would like anything changed. Thanks for taking a look.