TAESD previews with Z-Image-Turbo cause a massive increase in VRAM usage

I saw someone post something on Reddit about it, so I decided to test it for myself. I ran a basic workflow once to make sure everything was loaded and ready to go. After that, I tested TAESD vs Latent2RGB on a 1.2MP image(720x1600):

L2RGB: 10.8GB while diffusing
TAESD: 15.9GB while diffusing
Approximate a 47% increase in VRAM usage.

This is with the fp8_scaled version of the model

Launch script(using --fp32-vae because I get black images with fp16 on this AMD card):
```
set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
set MIOPEN_FIND_MODE=FAST
set MIOPEN_LOG_LEVEL=3
.\python_embeded\python.exe -s ComfyUI/main.py --disable-smart-memory --use-pytorch-cross-attention --fp32-vae
pause
```

Some other relevant information:
```
pytorch version: 2.10.0a0+rocm7.10.0a20251120
Set: torch.backends.cudnn.enabled = False for better AMD performance.
AMD arch: gfx1100
ROCm version: (7, 10)
Python version: 3.13.6 (tags/v3.13.6:4e66535, Aug 6 2025, 14:36:00) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.3.75
ComfyUI frontend version: 1.32.9
```

Any of the model loading logs show `full load: True`