[Evaluation] LM Harness refactor

[kaisopos]

Description

Refactoring our LM Harness integration code (specifically the mapping and setting of lm_eval.simple_evaluate's model_args, in preparation for resolving GH issue 1358 (support for Remote vLLM with LM Harness)

Why the refactor?
The model args are added all over the place (in model_params.py, in lm_harness.py at evaluate and _create_extra_lm_harness_model_params_for_vlm) and it's becoming increasingly buggy to track what has been added for 5 different engines and 2 modalities. Thus, aggregated everything in a single function, documented what params the 4-5 LH Harness engines currently support, the mapping from Oumi's var names, etc.

Related issues

Fixes OPE-940

Before submitting

• This PR only changes documentation. (You can ignore the following checks in that case)
• Did you read the contributor guideline Pull Request guidelines?
• Did you link the issue(s) related to this PR in the section above?
• Did you add / update tests where needed?

Reviewers

At least one review from a member of oumi-ai/oumi-staff is required.

[taenin]

Do we need to update any tests for this change?