[feat] Integrate LightEval, HuggingFace's new LLM evaluation framework

[supreme-gg-gg]

https://github.com/huggingface/lighteval

LightEval is literally amazing -- metrics, tasks, benchmarks, everything built in!!! This will be the replacement for our feature request on adding benchmark and expanding our evaluation capabilities #60

The design is that the user should have two ways to assess their model:

• Using the dataset it is trained on (use the test split of course), user preprocessed, compute metrics and stuff
• Use lighteval for task + benchmarks

I would say migrate from evaluate to lighteval since the former seems not as actively maintained as the latter!

[supreme-gg-gg]

My currently proposed design for this:

1. Add another mode called task evaluation -- these will use the tasks natively supported by LightEval and their associated metrics, very simple code just use the standard docs, only needs to ensure model compatibility with Unsloth + adapters

2. For the existing metrics based evaluation on user datasets, either replace the existing metrics with more powerful ones from LightEval or enhance our ones manually (e.g. utils for extracting responses) referencing LightEval. Note that LightEval docs is not up to date so source code reading is a must here (it's a very new library)