Add task variants replicating Llama 1 / 2 evaluation numbers

[haileyschoelkopf]

In some cases, Llama 1 (and 2, whose eval setups at times differ) paper results are not replicable by our implementations due to Meta’s custom undisclosed prompts or prepended task descriptions.

However, for some tasks like Triviaqa, we have successfully found setups / reverse engineered prompts. Where we have done this we should add documentation and variants of the tasks for ease of use.

[afcruzs]

@haileyschoelkopf Is this documented yet somewhere? Even if it's in an informal capacity (eg a branch, a Google doc etc)