Arm backend: Change node.name to buffer name when indexing state_dict (pytorch#12485)

Using node.name for indexing state_dict can result in the following:

`KeyError: 'b__tensor_constant1'`.

Using the following to fetch the buffer name is the correct way to do
it:
```
buffer_name = self.exported_program.graph_signature.inputs_to_buffers[
    node.name
]
```


Signed-off-by: Sebastian Larsson <[email protected]>

Author: Sebastian-Larsson

backends/arm/_passes/cast_int64_pass.py

@@ -47,7 +47,7 @@ def _to_int32(self, graph_module: torch.fx.GraphModule):
                 buffer_name = self.exported_program.graph_signature.inputs_to_buffers[
                     node.name
                 ]
-                buffer = self.exported_program.state_dict[node.name]
+                buffer = self.exported_program.state_dict[buffer_name]
                 self._assert_within_int32(buffer, node)
                 logger.warning(
                     f"Casting buffer {node.name} from torch.int64 to torch.int32"