Add a memory-efficient "chunked nll" loss type to SFTConfig and SFTTrainer .

Add a memory-efficient "chunked nll" loss type to SFTConfig and SFTTrainer .: a task in HF ML Tasksmith (Harbor dataset). loss type="chunked nll" computes the same next-token cross-entropy as the default "nll" path, but avoids materialising the full batch × seq len × vocab size logits tensor: