Nanotron greedy_until() fix

[vsabolcec]

Evaluating a nanotron checkpoint on a benchmark with a GreedyUntil request crashes with the following error message:

0: [rank0]:   File "/usr/local/lib/python3.10/dist-packages/lighteval/models/nanotron_model.py", line 1137, in greedy_until
0: [rank0]:     max_gen = max(item[1].generation_size for item in dataset)
0: [rank0]:   File "/usr/local/lib/python3.10/dist-packages/lighteval/models/nanotron_model.py", line 1137, in <genexpr>
0: [rank0]:     max_gen = max(item[1].generation_size for item in dataset)
0: [rank0]: TypeError: 'GreedyUntilRequest' object is not subscriptable

There is a bug in max_gen = max(item[1].generation_size for item in dataset) because dataset is GenerativeTaskDatasetNanotron whose __iter__() yields items directly without the index. This PR works around the issue in greedy_until() for the nanotron model.