Title: Native Normalized Attention Guidance node is ~2× slower than Kijai implementation

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

Native NAG performance should be comparable to (or faster than) Kijai’s implementation under equivalent conditions.

### Actual Behavior

Native NAG shows a consistent ~2× slowdown (total runtime and/or per-step time) compared to Kijai’s NAG.

### Steps to Reproduce

Steps to reproduce
1. Load the same checkpoint/model and VAE.
2. Use the same prompt/negative, seed.
3. Use identical sampler + scheduler, steps, CFG=1, batch size, and resolution.
4. Run the workflow with Kijai NAG enabled; record total time and/or per-step timing.
5. Swap to native NAG (keep all settings the same); record timing.
6. Compare results.

### Debug Logs

```powershell
100%|██████████| 4/4 [00:01<00:00, 2.89it/s] //kijai
100%|██████████| 4/4 [00:02<00:00, 1.84it/s] //native
```

### Other

The new native Normalized Attention Guidance (NAG) node in ComfyUI runs about 2× slower than the existing Kijai NAG implementation, using the same workflow and identical generation settings.