ZImageFunControlNet: Fix mask concatenation in --gpu-only (#11421)

This operation trades in latents which in --gpu-only may be out of the GPU
The two VAE results will follow the --gpu-only defined behaviour so follow
the inpaint image device when calculating the mask in this path.

Author: rattus128

comfy_extras/nodes_model_patch.py

@@ -348,7 +348,7 @@ def encode_latent_cond(self, control_image=None, inpaint_image=None):
             if self.mask is None:
                 mask_ = torch.zeros_like(inpaint_image_latent)[:, :1]
             else:
-                mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center")
+                mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True).to(device=inpaint_image_latent.device), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center")
 
             if latent_image is None:
                 latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(torch.ones_like(inpaint_image) * 0.5))