Allow passing a GenerationConfig for generative evals

[lewtun]

For some benchmarks, it can be preferable to generate responses via sampling with non-zero temperature. Although this affect reproducibility, it could be useful for power users who are constructing an internal benchmark with lighteval

One way to do this would be by allowing one to pass a generation_config field in the task metadata which is ultimately parsed by transformers.GenerationConfig (docs) before being passed to the generate() method

[clefourrier]

@lewtun can you check #428 to tell me if there are other params you'd want to see?