addmm_cuda not implemented for 'Float8_e4m3fn' Errors

### 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

Produce images with loras correctly

### Actual Behavior

Produces images without loras

### Steps to Reproduce

Since upgrading from ComfyUI v0.11.1, the **--fast** argument causes a regression when using the **z_image_turbo_fp8_e4m3fn** model. Specifically, the system logs errors (addmm_cuda" not implemented for 'Float8_e4m3fn) and fails to generate images when Loras are applied. The issue is bypassed by either removing the **--fast** flag or switching to the **z_image_turbo_bf16** model, though both workarounds result in a noticeable performance hit on my 4070 Ti. I am temporarily running v0.11.1 until this is resolved.

### Debug Logs

```powershell
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Requested to load ZImageTEModel_
Model ZImageTEModel_ prepared for dynamic VRAM loading. 7672MB Staged. 0 patches attached.
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16
Model ZImageTEModel_ prepared for dynamic VRAM loading. 7672MB Staged. 0 patches attached.
model weight dtype torch.float8_e4m3fn, manual cast: torch.bfloat16
model_type FLOW
Requested to load Lumina2
Model Lumina2 prepared for dynamic VRAM loading. 5869MB Staged. 180 patches attached.
0%| | 0/8 [00:00