[peft] fix: support ROCm TE grouped-linear fields

[Arist12]

Problem

ROCm Transformer Engine adds m_splits_tensor, actual_m_splits, and unpad_output to the private grouped-linear non_tensor_args contract. Bridge derives the installed field order but did not provide values for those names.

Change

Supply the public-wrapper defaults None, None, and False. Tuple construction still selects only fields exposed by the installed TE build, so CUDA contracts that omit them are unchanged.

Validation

• rebased onto current bridge@923842f5
• full diff passes repository pre-commit hooks
• focused contract test passes
• 4x MI350X GPT-OSS LoRA completes shared-outer and per-expert layouts

[Arist12]

hi @yushengsu-thu, could you help review this when you got a time? Thanks!