upgrade lm_eval to 0.4.5 (#6561)

Pull Request resolved: https://github.com/pytorch/executorch/pull/6533


We have been using a pretty old `lm_eval` version. This is blocking us from upgrading other libraries like `transformers` and blocking some others work. For example, https://github.com/pytorch/executorch/pull/6489.

In newer versions `lm_eval`, `pretrainedModel` becomes a required parameter. In 0.4.2, it defaults to `gpt2` if not provided. This PR upgrades our `lm_eval` version to the latest version 0.4.5 and set `pretrainedModel` to its original default value `gpt2`.

Differential Revision: [D65079913](https://our.internmc.facebook.com/intern/diff/D65079913/)
ghstack-source-id: 250754584

Co-authored-by: Lunwen He <lwhecser@gmail.com>