Clear VBAR state on deepclone to prevent shared C pointer race condition

When deepclone_multigpu deep-copies the model, ModelVBAR objects get copied
with their raw C pointers (_ptr). This causes two issues:
1. Double-free when the copied ModelVBAR is GC'd
2. Thread-safety crash: concurrent vbar_fault calls from multiple GPU threads
   corrupt the global linked list in model-vbar.c (no mutexes)

Fix: clear dynamic_vbars dict and remove _v attributes from all modules after
deepcopy, so each GPU clone gets fresh, independent VBARs during load().

Amp-Thread-ID: https://ampcode.com/threads/T-019d7168-759b-7090-8f31-c89dc6ac8d28
Co-authored-by: Amp <amp@ampcode.com>

comfy/model_patcher.py

@@ -414,6 +414,14 @@ class ModelPatcher:
             n.model = temp_model_patcher.model
         else:
             n.model = copy.deepcopy(n.model)
+        # Clear VBAR state so the clone gets fresh, device-specific VBARs during load().
+        # deep-copied ModelVBAR objects share raw C pointers with the original, which causes
+        # double-free and thread-safety issues (concurrent vbar_fault on shared global state).
+        if hasattr(n.model, "dynamic_vbars"):
+            n.model.dynamic_vbars = {}
+        for m in n.model.modules():
+            if hasattr(m, "_v"):
+                delattr(m, "_v")
         # unlike for normal clone, backup dicts that shared same ref should not;
         # otherwise, patchers that have deep copies of base models will erroneously influence each other.
         n.backup = copy.deepcopy(n.backup)