Adds a dummy/random model

[guipenedo]

Inspired by https://github.com/EleutherAI/lm-evaluation-harness/blob/main/lm_eval/models/dummy.py, this PR adds a random/dummy model that can be used to establish random baselines and to test/debug lighteval.
Can be run with

python run_evals_accelerate.py --model_args dummy ...

DummyModelConfig has a single option: seed. To pass a seed: --model_args dummy,seed=123. The seed is used to randomly generate logprobs

Two notes:

• had to change default values for truncated_tokens_count and padded_tokens_count otherwise evaluation_tracker.details_logger.aggregate crashes (https://github.com/huggingface/lighteval/blob/main/src/lighteval/logging/info_loggers.py#L420 and https://github.com/huggingface/lighteval/blob/main/src/lighteval/logging/info_loggers.py#L424)
• had to add an actual tokenizer (used gpt2) as even after overwriting tok_encode and tok_decode some other methods would still call the tokenizer directly

[NathanHB]

Great addition ! The code is very clean nothing to say. Could you add some doc in the README however ? :)