[BUG] Logs are missing some fields, truncation likely buggy

[clefourrier]

Describe the bug

With a model max length at 10K on ifeval, I'm getting truncations on the input (with no input stored in the logs): most likely an issue with the truncation.

I suspect it's also linked to the refacto on loglikelihood: the padding no longer works for accelerate models.

To Reproduce

lighteval accelerate 'model_name=Qwen/Qwen3-32B,model_parallel=True,max_length=16384,generation_parameters={temperature:0.6,top_p:0.9,top_k:40,min_p:0,max_new_tokens:16384}' 'extended|ifeval|0|0' --save-details --output-dir tmp

Logs will contain empty input fields in ModelResponse, with around 400 truncation

[clefourrier]

longest_context_continuation_size_in_split = len(tokenized_context) + split[0].generation_size should at least be longest_context_continuation_size_in_split = len(tokenized_context["input_ids"]) + split[0].generation_size

[clefourrier]

Padding is fixed now thanks to #901 but we still need to check the truncation numbers