fix: use non-deprecated is_autocast_enabled API and fix docstring

- torch.is_autocast_enabled() without device_type is deprecated in
  PyTorch 2.8, use torch.is_autocast_enabled(device_type) instead
- Update docstring to accurately describe root cause (upstream autocast
  leakage from nodes like SAM3, not fp8 specifically)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: deepme987

comfy/ldm/wan/model.py

@@ -175,10 +175,12 @@ def repeat_e(e, x):
 
 
 def _addcmul(x, y, z):
-    """torch.addcmul wrapper that disables autocast to avoid
-    'Unexpected floating ScalarType in at::autocast::prioritize' with fp8 weights."""
-    if torch.is_autocast_enabled():
-        with torch.autocast(device_type=x.device.type, enabled=False):
+    """torch.addcmul wrapper that disables autocast to prevent
+    'Unexpected floating ScalarType in at::autocast::prioritize' when
+    upstream nodes (e.g. SAM3) leave CUDA autocast enabled."""
+    device_type = x.device.type
+    if torch.is_autocast_enabled(device_type):
+        with torch.autocast(device_type=device_type, enabled=False):
             return torch.addcmul(x, y, z)
     return torch.addcmul(x, y, z)