Add experimental support for General Online Logit Distillation (GOLD) Trainer

[lukehinds]

https://huggingface.co/docs/trl/main/en/gold_trainer

Curate a GOLD dataset , for logit based teacher -> student distillation

from trl import GoldTrainer, GoldConfig
from datasets import load_dataset

dataset = load_dataset('json', data_files='training_data.jsonl')

config = GoldConfig(
    use_uld_loss=True,  # Universal Logit Distillation
    lmbda=0.5,          # Mix of student/teacher outputs
    beta=0.5            # Loss interpolation
)

trainer = GoldTrainer(
    model=student_model,
    teacher=teacher_model,
    args=config,
    train_dataset=dataset,  # DeepFabric data
)

trainer.train()