Consolidate PEFT support into Megatron-LM

[maanug-nv]

Goal

Move the PEFT support required by Megatron-Bridge into Megatron-LM so Bridge PEFT recipes can use megatron.training as the canonical training loop.

Scope

• Move performant LoRA, canonical LoRA, DoRA linear adapters, and the LoRALinear wrapper into Megatron Core.
• Apply adapters and freeze the base model before model wrapping in the Megatron-LM training loop.
• Add adapter-only checkpoint save and load support to megatron.training.checkpointing.
• Migrate Megatron-Bridge PEFT recipes to the Megatron-LM training loop.

Completion criteria

• Existing Megatron-Bridge PEFT APIs continue to work through the migration.
• Existing PEFT recipes retain their current behavior when run through the Megatron-LM training loop.
• Representative performant LoRA, canonical LoRA, and DoRA recipes preserve convergence and performance within agreed tolerances.
• Adapter-only checkpointing, adapter merging, and adapter export continue to work.