Make core aten exception ops modifiable at runtime.

Summary: Right now we run the ir verifier with a hardcoded set of whitelisted ops. This diff modifies the API to pass the user-defined ops to extend the exception list. See the attached task for more details.

Differential Revision: D75997936

Author: eigen-k

backends/cadence/aot/compiler.py

@@ -217,6 +217,7 @@ def _lower_ep_to_edge(
     expo_program: ExportedProgram,
     dump_graphs: bool = False,
     constant_methods: Optional[dict[str, object]] = None,
+    core_aten_exceptions: Optional[list[torch._ops.OpOverload]] = None,
 ) -> EdgeProgramManager:
     """
     Lower an ExportedProgram to an EdgeProgramManager (in edge IR).
@@ -237,7 +238,7 @@ def _lower_ep_to_edge(
                 torch.ops.aten.unfold.default,
                 torch.ops.aten.angle.default,
                 torch.ops.aten.rms_norm.default,
-            ],
+            ] + (core_aten_exceptions or []),
         ),
         constant_methods=constant_methods,
         preserve_ops=(torch.ops.aten.rms_norm.default,),
@@ -264,7 +265,8 @@ def export_to_edge(
     expo_program = trace(model, inputs)
 
     # Lower the model to edge IR.
-    edge_prog_manager = _lower_ep_to_edge(expo_program, dump_graphs, constant_methods)
+    edge_prog_manager = _lower_ep_to_edge(expo_program, dump_graphs, constant_methods,
+    core_aten_exceptions = [torch.ops.aten.linear.default])
 
     return edge_prog_manager