fix(qwen4): exclude PLE n-gram base state from PEFT saves

[0KEAHA]

When saving a Qwen4Exp LoRA adapter with peft_format=True, save_weights() calls
export_weights(..., _is_saving=True). The PLE export condition then evaluates
skip_ngram_state to False, so the adapter receives the base model's
layer_multipliers, ngram_heads_offsets, ngram_heads_vocab_sizes, and n-gram
embedding shards (plus the scale when emitted). These are not LoRA factors and
can cause adapter consumers to reject the checkpoint, as well as unnecessarily
exporting a very large PLE table.

Make PEFT exports skip the base n-gram state regardless of _is_saving.
Full-model saves still include it, and online full-model exports retain the
existing CPU-offload behavior. PLE projections continue through the existing
PEFT-aware _set_state_dict path. The adapter import fix added in #204 remains
unchanged.

Validation:

• Tested against upstream main at 35bc42c070e68dddf17c2043a0412a9be420cc95
  (checked on 2026-10-09); the same save issue is present in release/1.6 at
  73d6807a11acd8caf039a27459e3f4c377327a32 and swift-dev-v5 at
  99bf4022a3d20338d683b7b96513331bd32ab9bc.
• python -m pytest -q tests/test_qwen4_exp_peft_export.py: 5 tests and 22 subtests
  pass after the fix. The original source produces six expected failures in the adapter-save scenarios (and no test errors).
• Covers adapter export/save with CPU offload on/off and simulated PP1/PP2,
  adapter saving on a PP stage without the PLE layer, full-model saves, online
  full-model exports, and adapter/full-model imports.
• The tests compile the actual production method and buffer names from source;
  tensor operations, the table iterator, and PP collectives are mocked. They
  verify whether base buffers are exported and table export is scheduled.
• Flake8, isort, YAPF, and diff whitespace checks pass on the changed files.
• No multi-GPU checkpoint round trip or real vLLM loading has been run. This
  change addresses PLE base-state leakage; other LoRA module mapping issues
  remain outside its scope.