Use lighteval for benchmarking

[burtenshaw]

This PR implements a benchmarking with lighteval within the Colab & Runpod configuration.

There is still some remaining work to unify the code base:

• Implement averaging in the table
• Refactor table creation out of main.py The main features are tested:
• Multi GPU
• Preconfigured lighteval tasks like recommended_set.txt

[burtenshaw]

@mlabonne It would be really cool to contribute this. Let me know what you think about integrating it with the existing benchmarks.