nemotron_omni peft save skips all key conversion, adapters only work inside automodel

found while checking the omni adapters after #3595.

the nemotron_omni adapter converts keys by looking at how they start (language_model., sound_encoder., and so on). but peft keys start with base_model.model., so none of the checks match and the keys are saved as is, with no conversion at all.

that leaves three problems in the saved adapter:

• llm keys keep the internal model. naming instead of backbone.
• expert lora weights are saved as raw 3d tensors instead of the format peft expects
• sound encoder keys miss the encoder. part of their path

so the saved file uses names and formats that don't exist on the hf model, and hf peft can't load it. resuming inside automodel still works, but every load prints an "Unknown key prefix" warning for each tensor.

the fix: take the base_model.model. part off first, run the normal conversion, then put it back. the nemotron_v3 save side needs the same small change for its model. to backbone. rename.