Declare gradient accumulation loss scaling with loss_is_scaled_for_ga

[qgallouedec]

Uses Trainer.loss_is_scaled_for_ga (huggingface/transformers#49240) to say whether compute_loss already scales for gradient accumulation, instead of steering it through model_accepts_loss_kwargs and compute_loss_func.

• loss_is_scaled_for_ga = False (Trainer divides by GA steps): DPO, KTO, RLOO, Reward, CPO, ORPO. Replaces self.model_accepts_loss_kwargs = False.
• loss_is_scaled_for_ga = True (trainer normalizes over the accumulated batch itself): GRPO, Distillation, AsyncGRPO, AsyncDistillation, SDFT, SDPO, SSD. Replaces compute_loss_func="non-None value to disable scaling" plus self.model_accepts_loss_kwargs = False.
• SFT keeps the default None.

For transformers < 5.19, _BaseTrainer.__init__ maps the flag back to the two attributes, so behaviour there is unchanged. That block goes away when the floor reaches 5.19.

Step-1 grad_norm for DPO, Reward and GRPO, batch 4 x GA 1 and batch 2 x GA 2: identical to main on transformers 5.18 and on transformers with #49240.

Merge after huggingface/transformers#49240.