Arm backend: Update partitioner de-tagging iteration order (#10813)

When de-tagging tagged/partitioned nodes, the order in which we iterate
the nodes will matter as each node will make a decision based on if the
parent node is partitoned or has a compatible dtype. This change makes
sure we iterate the nodes in the order of the original graph_module.

Signed-off-by: Oscar Andersson <[email protected]>

Author: oscarandersson8218

backends/arm/tosa_partitioner.py

@@ -96,7 +96,9 @@ def is_partitioned(node: torch.fx.Node, tag=tag) -> bool:
 
             # De-tag outmost q-nodes upwards and dq-nodes downwards.
             # De-tag if at least one input/ output is not part of partition.
-            for node in partition.nodes:
+            for node in exported_program.graph_module.graph.nodes:
+                if not is_partitioned(node):
+                    continue
                 if is_quant_node(node):
                     for input in node.all_input_nodes:
                         if not is_partitioned(input):