Fix Genesis RuntimeError: don't move simulator tensors to GPU

Genesis uses backend=gs.cpu — all simulation runs on CPU. Moving
simulator tensors to CUDA caused two bugs:
1. scene.reset() raises "Tensor subclass already associated" because
Genesis internally tries to re-wrap CPU tensors that conflict with
our moved CUDA tensors.
2. state.pos from solver.get_state() is always on CPU, so
particles_now - init_particles_gpu would be a CPU/CUDA mismatch.

Fix: remove simulator.move_to_device() calls. Only the video generator
pipeline (transformer, VAE, text encoder) and case data (latents,
embeddings) are moved to GPU. Simulation and PyTorch3D rendering stay
on CPU throughout.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -489,11 +489,10 @@ def do_generate(case_name, prompt, d0, s0, d1, s1, d2, s2):
         return
 
     # Transfer all CPU-resident state to GPU for this generation session.
+    # NOTE: simulators are NOT moved to GPU — Genesis uses backend=gs.cpu and
+    # simulation tensors must remain on CPU alongside Genesis internal state.
     video_generator.move_pipeline_to_device("cuda")
     video_generator.move_case_data_to_device("cuda")
-    for _b in cases.values():
-        if _b.simulator is not None:
-            _b.simulator.move_to_device("cuda")
 
     bundle = cases[case_name]
 
@@ -686,9 +685,6 @@ def do_generate(case_name, prompt, d0, s0, d1, s1, d2, s2):
         if video_generator is not None:
             video_generator.move_pipeline_to_device("cpu")
             video_generator.move_case_data_to_device("cpu")
-        for _b in cases.values():
-            if _b is not None and _b.simulator is not None:
-                _b.simulator.move_to_device("cpu")
         torch.cuda.empty_cache()
         _is_generating = False