Generated image is different in FLUX right after commit ab130001a8b966ed788f7436aa3b689d038e42a3 ("Do RMSNorm in native type")

### Expected Behavior

This image

### Actual Behavior

This image

### Steps to Reproduce

-Make sure you're in commit ca4b8f30e0bf40cf58dcb3f3e6118832a60348c8
-Now load the attached workflow (it's the default FLUX1.dev example, only changed to FP8, and PREVIEW image instead of SAVE image node) and generate the image
[flux dev example.json](https://github.com/user-attachments/files/16914580/flux.dev.example.json)

-Save the image

-Now switch to commit ab130001a8b966ed788f7436aa3b689d038e42a3 (the one immediately after ca4b8f30e0bf40cf58dcb3f3e6118832a60348c8), and generate the image using the same workflow.

-Now compare the image with the previous.

### Debug Logs

```powershell
not applicable
```

### Other

Note:

-this workflow doesn't have any custom nodes.

-I've generated the image more than once in each commit, and for a same commit, the result is deterministic

-The problem also happens to Flux Dev Q8 GUFF model (this one needs a custom node)

**Launch arguments:**
--windows-standalone-build --disable-auto-launch --disable-xformers --temp-directory "%tmp%\Comfyui" --preview-method latent2rgb

**Environment details:**

Total VRAM 16376 MB, total RAM 131022 MB
pytorch version: 2.2.2+cu121
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4080 : cudaMallocAsync
Using pytorch cross attention

**Computer:**
Windows 11 Pro
Using Intel i9-9900K
128GB RAM
RTX 4080 (16GB RAM)