significant increase in processing time since update 0.4.0

### 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

I am running a simple flux.1(Dev) Workflow, which took under 2 mins yesterday. Today I can't wait it out, after 15 min generation time i simply give up! I am running Comfy on a Laptop: Windows 11(25H2), 64Gig Ram, a RTX 5080 (newest drivers) AMD Ryzen 9 9955HX with integr. AMD Graphics (2GiG). My suspicion is that the darn program uses my built in graphics! I WANT VER. 0.3.76 BACK!!!! I hate the missing of Comfy Manager, the absolute reluctance to incorporate ANY custom nodes and the f...-up settings! PLEASE ROLL BACK!!!

### Actual Behavior

Most of my workflows are tecnically working again after 13 Hours of cussing, craying and throwing things around... just to find out, that my image generation speed is non existant (Flux.1(Dev) took 17 min for about 20 per cent in KSampler on the standart workflow with NO CUSTOM NODES!!!)

### Steps to Reproduce

[flux_dev_full_text_to_image.json](https://github.com/user-attachments/files/24110424/flux_dev_full_text_to_image.json)

### Debug Logs

```powershell
The darn thing has no errors, its technically working, just f...in slow!!!
```

### Other

_No response_

Start by reproducing the slowdown with the attached flux_dev_full_text_to_image.json workflow, using the reported Windows 11 and RTX 5080 setup. Compare generation time and device usage with the reported under-2-minute baseline, then inspect the KSampler stage and available debug output. Done means the regression is isolated and generation returns to the expected performance, or the issue has actionable diagnostics.

Written by the indexing model from the issue text.