fix: MoE lora adapter layout

[akoumpa]

What does this PR do ?

TestMoELoRASaveRestoreMergeHF covers -- the exact workflow from issue #1226:

1. test_peft_loads_adapter -- Grouped LoRA tensors are converted to per-expert HF format, saved as adapter_model.safetensors + adapter_config.json, and PeftModel.from_pretrained() loads them without error.
2. test_merge_changes_all_expert_weights -- After merge_and_unload(), every expert's gate_proj, up_proj, and down_proj weights differ from the base (all 24 = 2 layers x 4 experts x 3 projections).
3. test_merge_values_are_numerically_correct -- The merged weight values exactly satisfy W_merged = W_base + (lora_B @ lora_A) * (alpha / r), verifying the LoRA delta is correctly applied.
4. test_merge_removes_lora_params -- After merge, no lora_ parameters or modules remain in the model.

Changelog

• Add specific line by line info of high level changes in this PR.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to # (issue)

[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.

[akoumpa]

/ok to test 8915726

[akoumpa]

/ok to test 750f55a

[akoumpa]

/ok to test dad2a6f

[akoumpa]

/ok to test 90eae91

[HuiyingLi]

lgtm.