Align Liger loss naming across experimental trainers

[albertvillanova]

Align Liger loss naming across experimental trainers.

Follow-up to:

• Align Liger loss naming across core trainers #6243

This PR standardizes the naming of the Liger loss instance variable across multiple experimental trainer classes. The variable previously named liger_jsd_loss is now consistently renamed to liger_loss wherever it is initialized and used. This improves code clarity and consistency across the codebase.

Changes

Variable renaming for consistency:

• Renamed the liger_jsd_loss instance variable to liger_loss in the constructors of the following trainer classes: DistillationTrainer, GKDTrainer, GOLDTrainer, SDFTTrainer, and SDPOTrainer (trl/experimental/distillation/distillation_trainer.py, trl/experimental/gkd/gkd_trainer.py, trl/experimental/gold/gold_trainer.py, trl/experimental/sdft/sdft_trainer.py, trl/experimental/sdpo/sdpo_trainer.py).

• Updated all usages of the renamed variable from self.liger_jsd_loss to self.liger_loss in the corresponding loss computation methods in each trainer.

---

Note

Low Risk
Identifier-only refactor with no training or loss behavior changes.

Overview
Renames the fused Liger JSD loss instance from liger_jsd_loss to liger_loss in five experimental trainers: DistillationTrainer, GKDTrainer, GOLDTrainer, SDFTTrainer, and SDPOTrainer.

Initialization and loss forward calls are updated in each file; LigerFusedLinearJSDLoss construction and arguments are unchanged. This aligns naming with follow-up to #6243 so the attribute name is consistent across distillation-style trainers.

^{Reviewed by Cursor Bugbot for commit 416fbd2. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.