[ONNX] Enable experimental exporter logic to dynamo_export and support refine dynamic_shapes (#134976)

(1) Enable experimental exporter logic to dynamo_export
(2) Refine dynamic shapes and retry export in export strategies
(3) Delete `torch_export_graph_extractor` and use the new export logic
(4) Disable ExportedProgram test in `test_fx_onnx_with_onnxruntime.py`, as ONNXProgram is different now.

Fixes https://github.com/pytorch/pytorch/issues/126479
Fixes #135183
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134976
Approved by: https://github.com/justinchuby

test/onnx/pytorch_test_common.py

@@ -346,6 +346,28 @@ def skipDtypeChecking(func):
     return wrapper
 
 
+def skip_if_fake_model_and_inititalizer(reason: Optional[str] = None):
+    """skip test with models using ExportedProgram as input.
+
+    Args:
+        reason: The reason for skip the ONNX export test.
+
+    Returns:
+        A decorator for skip tests.
+    """
+
+    def skip_dec(func):
+        @functools.wraps(func)
+        def wrapper(self, *args, **kwargs):
+            if kwargs["use_fake_mode"] and kwargs["include_initializer"]:
+                return unittest.SkipTest(reason)
+            return func(self, *args, **kwargs)
+
+        return wrapper
+
+    return skip_dec
+
+
 def xfail_if_model_type_is_exportedprogram(
     error_message: str, reason: Optional[str] = None
 ):