fix(inpainting): keep LanPaint sigma on the input device

[acesof]

Problem

Inpainting crashes on any non-CPU device that isn't the default, e.g. Apple Silicon:

RuntimeError: Expected all tensors to be on the same device,
but found at least two devices, mps:0 and cpu!

Traceback:

File "models/qwen/pipeline_qwenimage.py", line 979, in __call__
  latents = lanpaint_proc(denoise, cfg_predictions, true_cfg_scale, 1., latents,
                          original_image_latents, randn, t/1000, image_mask_latents, ...)
File "shared/inpainting/lanpaint.py", line 87, in __call__
  out = self.LanPaint(...)
File "shared/inpainting/lanpaint.py", line 142, in LanPaint
  noisy_image = self.latent_image * (1.0 - sigma) + self.noise * sigma

Cause

The pipeline passes t/1000, which is a 0-dim tensor, so this branch fires:

if len(sigma.shape) == 0:
    sigma = torch.tensor([sigma.item()])   # no device= -> always CPU

torch.tensor(...) without device= allocates on CPU regardless of where
sigma came from. Line 142 then multiplies device-resident latents by that CPU
tensor.

Note this is a shape-(1,) tensor, not a 0-dim one, so it does not qualify for
PyTorch's CPU-scalar promotion and the op genuinely fails.

Fix

-            sigma = torch.tensor([sigma.item()])
+            sigma = torch.tensor([sigma.item()], device=sigma.device)

I used device= rather than sigma.unsqueeze(0) deliberately:
torch.tensor([x.item()]) infers float32 regardless of the input dtype, so
adding device= changes placement only and leaves numerics identical.
unsqueeze(0) would preserve bf16 and silently change the precision of the
sigma schedule downstream.

shared/inpainting/ was checked for the same pattern — this is the only
device-less tensor construction. utils.py:270,280,293 already pass
device=y0.device.

Verification

Apple M4 Pro / MPS, torch 2.13.0, qwen_image_20B, 512x512,
video_prompt_type: "VA", model_mode: 2, 3 steps.

Before: crashes at the first LanPaint call.
After: completes in ~1m07s, exit 0, no errors in the log.

To confirm the mask is genuinely respected rather than the run merely not
crashing, mean absolute pixel difference from the source image:

region	mean abs diff
inside mask	51.81
outside mask	0.96

The masked region is regenerated; the rest is preserved (0.96 is JPEG noise).

Not exercised: the IS_FLUX / IS_FLOW branches and Flux/Wan inpainting. The
changed line runs before that branch, so they are covered by the same fix, but I
only ran the Qwen path.

🤖 Generated with Claude Code

[acesof]

Reopening, closed by mistake.