Cortex-M backend: Ensure no get_attr nodes appear after lowering (#18538)

Adds an additional run of the lift_constant_tensor_pass after all passes
in CortexMPassManager has been run, previously the serializer would
complain in some cases with ETRecords enabled, see #18422

Signed-off-by: Adrian Lundell <adrian.lundell@arm.com>

Author: AdrianLundell

backends/cortex_m/passes/cortex_m_pass_manager.py

@@ -17,7 +17,7 @@
 )
 from executorch.exir.pass_base import ExportPass
 from executorch.exir.pass_manager import PassManager
-from executorch.exir.program._program import _transform
+from executorch.exir.program._program import _transform, lift_constant_tensor_pass
 from torch.export import ExportedProgram
 from torch.fx.passes.infra.pass_base import PassResult
 
@@ -81,4 +81,8 @@ def transform(self) -> ExportedProgram:
                 transform_pass = pass_cls()
             pass_callable = cast(Callable[[Module], PassResult], transform_pass)
             ep = _transform(ep, pass_callable)
+
+        # All constant tensors should be lifted to buffers at this point, re-run
+        # lift_constant_tensor_pass in case new ones have been introduced by the passes above.
+        ep = lift_constant_tensor_pass(ep)
         return ep