Add tasks for performance on long context lengths

[Brian Vaughan]

There are a couple of papers I see with benchmarks for really long context lengths that don't seem to be available in lm-evaluation-harness. It would be great to have one of these or something similar for measuring ability to extract information from long context windows, important for RAG.

• LongBench: https://arxiv.org/abs/2308.14508
• ∞ Bench: https://arxiv.org/abs/2402.13718
• Are there any others that might be better?

[Hailey Schoelkopf]

Needle-in-a-haystack might also be a nice-to-have though I think more difficult / "natural" long-context evals should be prioritized.