fix(peft): patch the LayerNorm copies, not the ModulesToSaveWrapper itself

[yurekami]

Root cause

The PEFT branch of _patch_layer_norm_module (src/liger_kernel/transformers/monkey_patch.py:103) diverges from its _patch_rms_norm_module sibling in three ways, all defects:

1. It binds LigerLayerNorm.forward/extra_repr onto the ModulesToSaveWrapper itself. The instance-level forward shadows PEFT's _forward_wrapped dispatch (which routes to the trainable modules_to_save.default copy when adapters are enabled, and to original_module under disable_adapters), silently breaking PEFT semantics. The RMSNorm branch from the same PR correctly leaves the wrapper untouched.
2. hidden_size is set on the wrapper and (twice) on original_module, but never on modules_to_save.default — yet extra_repr (which reads self.hidden_size) is bound to it, so repr()/print() of a patched model raises AttributeError: 'LayerNorm' object has no attribute 'hidden_size'.
3. The duplicated original_module.hidden_size assignment is dead weight.

The branch is reached whenever apply_liger_kernel_to_*(model=...) patches a LayerNorm wrapped by a PEFT config's modules_to_save (e.g. vision-tower layernorms in mllama/InternVL/SmolVLM-style setups). It was introduced by #632 (which fixed #631 for the RMSNorm path) and has zero test coverage — grep -c peft test/transformers/test_monkey_patch.py on main is 0.

Fix

Mirror _patch_rms_norm_module: patch only modules_to_save.default and original_module, set hidden_size on both, and leave the wrapper's forward alone so PEFT's adapter dispatch stays intact. Net −3 lines in src. Adds a regression test covering the PEFT wrapper path, including the invariant that "forward" not in wrapper.__dict__.

Verification

Executed on CPU (WSL, torch 2.13, transformers 4.56.2, peft 0.20.0):

• Red: on main, the new test fails with AttributeError: 'LayerNorm' object has no attribute 'hidden_size' on print(wrapper), and the wrapper's __dict__ carries the shadowing forward.
• Green: with the fix, full test/transformers/test_monkey_patch.py passes — 44 passed, 19 skipped, 0 failed. ruff check and ruff format clean.

Disclosure: the GPU kernel suite (make test) and convergence tests were not run — no GPU on this box. The change touches only patch-time attribute/binding structure, no kernel code.