Apply PEFT adapters in the Megatron-LM training loop

[maanug-nv]

Parent initiative: #7267

Work

Add a pre-wrap PEFT hook to the Megatron-LM training path. The hook must apply configured adapters and freeze the base model before distributed model wrapping.

Done when

• megatron.training can apply the migrated PEFT adapters during model setup.
• Adapter application happens before DDP or FSDP wraps the model.
• Base-model parameters are frozen while adapter parameters remain trainable.
• The path supports performant LoRA, canonical LoRA, and DoRA.
• Tests verify adapter placement, parameter freezing, and ordering relative to model wrapping.