Add CLI to evaluate models, list supported tasks etc

[lewtun]

If would be nice if one could launch evaluation jobs directly from the command line with something like:

lighteval --tasks="lighteval|hellaswag|5|1" --output_dir "/scratch/evals" --model_args "pretrained=gpt2

By default we could look for the default accelerate config, but allow users to override this if needed with

lighteval --accelerate_config=path/to/accelerate/config --tasks="lighteval|hellaswag|5|1" --output_dir "/scratch/evals" --model_args "pretrained=gpt2

It would also be nice if the CLI would produce a list of supported tasks with something like

lighteval --list-tasks

[clefourrier]

We now have a better CLI and can show supported task. #228 will soon allow to display samples too, so I'm closing :)