Add Entropy Adaptive Fine Tuning to SFT Trainer

[electroglyph]

from the paper: https://huggingface.co/papers/2601.02151

the implementation is 100% based on (and credited to):
hiyouga/LlamaFactory@main...ymxyll:LlamaFactory-EAFT:feature/eaft

i added a test =)

closes #4795

use like so:

trainer = SFTTrainer(
    args = SFTConfig(
        loss_type = "eaft",
        eaft_alpha = 1.0, # default
    ),
)

[qgallouedec]

Thanks! Can you add a short subsection here: https://github.com/huggingface/trl/blob/main/docs/source/paper_index.md#supervised-fine-tuning

[electroglyph]

Thanks! Can you add a short subsection here: https://github.com/huggingface/trl/blob/main/docs/source/paper_index.md#supervised-fine-tuning

just letting you know i added the section