Fix global LayerNorm mutation in model patching

[vedavyas2727]

Summary

Fixes #315.

Prevent class-level LayerNorm patching for Mllama, Gemma3/PaliGemma (SigLIP), InternVL, and SmolVLM from replacing torch.nn.LayerNorm process-wide. Each Transformers modeling module now receives a private torch.nn namespace whose LayerNorm alone points to LigerLayerNorm.

Details

The affected modeling modules import torch.nn as nn, so assigning modeling_module.nn.LayerNorm previously mutated the shared torch.nn module. That could unexpectedly change LayerNorm construction in unrelated models and application code after enabling Liger.

The helper clones the modeling module's nn namespace before overriding LayerNorm. InternVL also caches nn.LayerNorm in NORM2FN, so that model-local factory is updated explicitly to preserve Liger patching for newly constructed vision layers.

No benchmark is applicable because this changes patch isolation only; it does not alter a kernel or runtime computation path.

Testing Done

• Hardware Type: Windows CPU (torch 2.12.0+cpu)
• run make test to ensure correctness (CUDA/Triton kernel suite unavailable on this CPU-only host)
• run checkstyle-equivalent ruff check --output-format=concise . and ruff format --check --diff .
• run make test-convergence to ensure convergence (requires an accelerator)
• transformers 5.12.0: pytest test/transformers/test_monkey_patch.py -x -q — 68 passed
• Minimum supported transformers 4.52.0: targeted class-level regression suite — 6 passed
• Independent code review completed; its InternVL cached-factory finding was addressed and the final review found no remaining significant issues.