[Megatron] Fix reference adapter loading for LoRA RLHF

[taking-lying-flat]

Summary

• load --ref_adapters into the policy model chunks that are actually used for adapter-based reference forwards
• load --mcore_ref_adapter into ref_adapter by mapping source default lookup keys to the target reference slot
• support safetensors adapter loading for lora_llm in both policy and reference slots
• normalize one newly added Qwen f-string for the Python 3.10 lint runner

Root cause

MegatronRLHFTrainer.prepare_model() initialized self.ref_models as an empty list for LoRA training, but passed that list to bridge.load_weights() for --ref_adapters. The bridge therefore opened the checkpoint but converted zero model chunks. Reference forwards later switched self.unwrapped_models to ref_adapter, so DPO/KTO/GRPO used the freshly initialized adapter instead of the requested checkpoint without raising an error.

The MCore checkpoint path separately used the default adapter_name and retained every PEFT key without remapping its checkpoint lookup key. A normal policy LoRA checkpoint therefore matched and overwrote the current model default adapter instead of initializing ref_adapter.

Impact

Megatron RLHF runs using LoRA plus either reference-adapter format now load the requested weights into the adapter used for reference log-probability computation without overwriting the policy adapter. lora_llm no longer silently skips safetensors policy or reference adapters.