Add Uncheatable Eval

[ziqing-huang]

This PR add Uncheatable Eval to the task list. It evaluats LLMs on fresh, real-time internet data so the tests are leak-proof. Unlike static benchmarks that risk pretraining contamination, Uncheatable Eval uses fresh, real-time internet data, ensuring the model has never seen the test samples. This makes the evaluation leak-proof and gives a more accurate measure of true generalization, especially for base models. Integrating it into LM Eval Harness lets us run these contamination-resistant tests in a standard, unified interface.

[CLAassistant]

All committers have signed the CLA.