fix(qwen2_5_omni): place thinker prefix inside base_model.model on PEFT saves

[taran-dev4u]

Fixes #3655.

Problem

When saving LoRA adapters for \qwen2_5_omni, parameter keys already start with the PEFT prefix \�ase_model.model..
\Qwen2_5OmniStateDictAdapter.to_hf\ previously prepended \ hinker.\ directly to the entire key, producing:
\\ ext
thinker.base_model.model.model.layers.0.self_attn.q_proj.lora_A.weight
\
instead of the expected Hugging Face PEFT format:
\\ ext
base_model.model.thinker.model.layers.0.self_attn.q_proj.lora_A.weight
\
This prevented external Hugging Face PEFT from loading saved adapters and broke bidirectional conversion in \rom_hf.

Solution

1. In \ o_hf: Check for \�ase_model.model.\ prefix and place \ hinker.\ inside it (\�ase_model.model.thinker...).
2. In \rom_hf: Strip \ hinker.\ when nested inside \�ase_model.model., preserving the PEFT prefix.
3. In \convert_single_tensor_to_hf: Apply identical prefix nesting logic for single tensor conversions.
4. Added comprehensive unit tests for PEFT LoRA key conversion, bidirectional round-trip serialization, and single-tensor export.

Testing

• Passed 13/13 unit tests in \ ests/unit_tests/models/qwen2_5_omni/test_qwen2_5_omni_state_dict_adapter.py.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.