Support preserving calling convention to some modules.

Summary:
X-link: pytorch/pytorch#106798

APS use this feature to swap out some submodules after unflattening.

Reviewed By: kit1980, tugsbayasgalan

Differential Revision: D48154341

fbshipit-source-id: 687393582596d679a51f236ae1665a5e07cdaadc

Author: zhxchen17

exir/capture/_capture.py

@@ -212,6 +212,7 @@ def convert_to_fake(x):
         {},
         {},
         [],
+        [],
     )
     return ExirExportedProgram(ep, False)
 

exir/lowered_backend_module.py

@@ -233,6 +233,7 @@ def create_exported_program_from_submodule(
         # TODO(T159524653): fill in range and equality constraints
         range_constraints={},
         equality_constraints=[],
+        module_call_graph=[],
     )
 
 

exir/serde/serialize.py

@@ -42,7 +42,7 @@ class GraphModuleSerializer(export_serialize.GraphModuleSerializer):
     def __init__(
         self, graph_signature: ep.ExportGraphSignature, call_spec: ep.CallSpec
     ) -> None:
-        super().__init__(graph_signature, call_spec)
+        super().__init__(graph_signature, call_spec, [])
         self.state_dict: Dict[str, torch.Tensor] = {}  # TODO(T157676982)
 
     def serialize_operator(
@@ -574,6 +574,7 @@ def deserialize(
             graph_module,
             sig,
             call_spec,
+            module_call_graph,
             symbol_name_to_symbol,
         ) = GraphModuleDeserializer(state_dict).deserialize(
             serialized_exported_program.graph_module,
@@ -603,6 +604,7 @@ def deserialize(
             {},  # TODO(T157676982)
             range_constraints,
             equality_constraints,
+            [],
         )