Arm backend: Fix for tensor capture in quantization assessment. (#24279)

Fix bug in the assessment tool.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Signed-off-by: Elena Zhelezina <elena.zhelezina@arm.com>

Author: wwwind

backends/arm/vgf/quantization_assessment.py

@@ -32,6 +32,8 @@
 from torch.export import ExportedProgram
 from torch.fx import GraphModule, Interpreter, Node
 
+from torch.utils._pytree import tree_map
+
 
 GraphInput = ExportedProgram | GraphModule
 
@@ -187,8 +189,13 @@ def __init__(self, module: GraphModule, capture_node_names: set[str]) -> None:
 
     def run_node(self, node: Node) -> Any:
         value = super().run_node(node)
+
         if node.name in self._capture_node_names:
-            self.captured[node.name] = value
+            self.captured[node.name] = tree_map(
+                lambda x: (x.detach().clone() if isinstance(x, torch.Tensor) else x),
+                value,
+            )
+
         return value