[Modular Converter] Strip @use_kernelized_func(fn) when fn lacks @use_kernel_func_from_hub

[kpal002]

What this PR does

Fixes the root cause of the use_kernels=True crash on Qwen3Next and Qwen3.5 (issue #46399).

Root cause

replace_class_node() falls back to the parent class decorators when a modular subclass declares none. For Qwen3NextAttention(Qwen3MoeAttention) and Qwen3_5Attention, this means @use_kernelized_func(apply_rotary_pos_emb) is inherited from Qwen3MoeAttention — but the generated file defines apply_rotary_pos_emb as a partial-RoPE variant without @use_kernel_func_from_hub. When kernelize() runs, attach_hidden_kernels() tries to register that plain function as an nn.Module and raises:

ValueError: Attempted to register a kernel for apply_rotary_pos_emb,
but it was not a torch.nn.Module.

Fix

Add strip_orphaned_kernelized_func_decorators(), called in create_modules() before get_needed_imports().

For each generated file's body dict it:

1. Collects all module-level function names and identifies which have @use_kernel_func_from_hub
2. For any class with @use_kernelized_func(fn) where fn is defined locally without @use_kernel_func_from_hub, drops that decorator

Calling it before get_needed_imports means the existing LibCST scope analysis also removes the now-unused use_kernelized_func import automatically — no separate cleanup needed.

What is NOT changed

• No model files are touched here. The regeneration of modeling_qwen3_next.py and modeling_qwen3_5.py (using this fixed converter) is in a follow-up PR to keep the diff reviewable.
• No kernel is added for partial RoPE. That belongs in a future PR once kernels-community#919 ships.

PR series for #46399

PR	Status	Description
#46435	Open	Regression tests (currently failing — will pass after this)
This PR	Open	Converter fix
PR 3	Pending	Regenerate modeling_qwen3_next.py and modeling_qwen3_5.py

cc @vasqu @ArthurZucker @MekkCyber @drbh

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Was this discussed/approved via a Github issue or the forum? Yes — Qwen3.5: use_kernels=True crashes with ValueError on apply_rotary_pos_emb registration - missing decorator #46399
• Did you update the documentation with your relevant changes?
• Did you write any new necessary tests? Regression tests in [Tests] Add regression tests for use_kernels=True crash on Qwen3Next and Qwen3.5 #46435; unit tests for this function in PR 3.