upgrade lm_eval to 0.4.5

[helunwencser]

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, #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.

Test Plan:
Run eval before and after this PR. Make sure the perplexity number stays around the same.

Differential Revision: D65162734

[helunwencser]

Stack from ghstack (oldest at bottom):

• -> upgrade lm_eval to 0.4.5 #6555

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6555

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 11a650e with merge base 2c32bf3 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[helunwencser]

@helunwencser has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.