fix: PTD hop schedule keeps the sampler's device and dtype

The schedule handed to step-aware code (SLA, H3's head bank) was a CPU copy while the call's sigma sits on the GPU, which fails with "Expected all tensors to be on the same device". It now matches the schedule the sampler was given. Test uses the meta device as a stand-in for CUDA.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0164nckPWA2L2ugVqiDcXzc3

modules/sampling/ptd/__init__.py

@@ -102,13 +102,16 @@ def _swap(controller):
         except Exception:      # noqa: BLE001 -- only the progress bar and previews are lost
             pass
         # Step-aware modules (SLA, momentum, velocity bias) read the schedule from here: give them the real one.
+        # On the device and dtype the sampler was given: ComfyUI's own schedule sits on the GPU, and a model that
+        # compares it with its call's sigma (H3's head bank, SLA) fails on a CPU copy.
+        grid = controller.grid.to(device=sigmas.device, dtype=sigmas.dtype)
         options = extra_args.get("model_options", {}).get("transformer_options")
         if options is not None:
-            options["sample_sigmas"] = controller.grid.float()
+            options["sample_sigmas"] = grid
         log.once("ptd:swap", log.INFO, "FunPack PTD",
                  f"sampling {controller.hops} hops from {controller.adapter.label}; the chosen sampler and steps "
                  f"({len(sigmas) - 1}) are not used")
-        return ptd.sample(guider, controller.grid.float(), extra_args, callback, noise, latent_image,
+        return ptd.sample(guider, grid, extra_args, callback, noise, latent_image,
                           denoise_mask, disable_pbar)
 
     return run