Fix conv2d bias INT48 type hint removed by fusing placeholder nods (#15257)

Summary:

Fix conv2d bias INT48 type hint removed by fusing placeholder nods

Reviewed By: Ninja91, digantdesai

Differential Revision: D84948170

Author: 3l1

backends/arm/_passes/fuse_equal_placeholders_pass.py

@@ -83,6 +83,13 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
                     rep_tensor,
                     common_persistent,
                 )
+                
+                # TBD: Find a principled way to merge node.meta across all fused node
+                # For now, i specifically transfer over the TosaSpecialDtype.meta_key() of the rep_node
+                if TosaSpecialDtype.meta_key() in rep_node.meta:
+                    common_node.meta[TosaSpecialDtype.meta_key()] = rep_node.meta[
+                        TosaSpecialDtype.meta_key()
+                    ]
 
             # Replace uses and delete duplicates
             for node, _ in nodes_tensors:

backends/arm/test/targets.bzl

@@ -18,6 +18,7 @@ def define_arm_tests():
         "ops/test_addmm.py",
         "ops/test_avg_pool2d.py",
         "ops/test_cat.py",
+        "ops/test_conv2d.py",
         "ops/test_linear.py", 
         "ops/test_mul.py",
         "ops/test_permute.py",