fix steering baseline

[xwh12345-user]

This PR fixes ori_generate() / ori_vllm_generate() so original predictions are generated without active interventions.

What Changed

• Added shared generation-state helpers in steer/models/model_wrapper.py:

  • _save_and_clear_generation_state()
  • _restore_generation_state()

• ori_generate() now temporarily clears and restores both:

  • add_activations_dict
  • intervention_dict

• ori_vllm_generate() now applies the same cleanup/restore logic.

• GPTWrapper.ori_generate() was also updated to use the shared helper logic.

This ensures orig_pred is generated from the base model state, without being affected by RePS/SFT/SPILT-style interventions stored
in intervention_dict.

What Was Tested

• Static checks:

  • python -m py_compile steer/models/model_wrapper.py
  • git diff --check

• Unit-style state restoration test:

  • Verified add_activations_dict and intervention_dict are both cleared before generation.
  • Verified both are restored after generation.
  • Verified intervention objects are restored as the same original objects.

• Real-model tests on GPU 0:

  • /disk4/xuweihong/models/Qwen2.5-7B-Instruct
  • /disk4/xuweihong/models/Qwen3-VL-4B-Instruct

Results

• Qwen2.5-7B-Instruct

  • Baseline ori_generate(): Paris
  • Normal generate() with intervention: London
  • Fixed ori_generate() with intervention: Paris
  • Intervention restored after generation: yes

• Qwen3-VL-4B-Instruct

  • Normal generate() was affected by intervention.
  • Fixed ori_generate() matched the baseline output.
  • Intervention restored after generation: yes