Fix S367052 to unblock ICVR MC3 (pytorch#109853)

Summary: Somehow "getitem" started to get Tensor starting from ads_ranking:996 and broke SDD pipelining FX-transformer. We need to skip the Tensor node in annotation.

Test Plan:
N4326037

# Before
 {F1099052907}

# With this diff

 {F1099052270}

Differential Revision: D49528046

Pull Request resolved: pytorch#109853
Approved by: https://github.com/jackiexu1992, https://github.com/lanza, https://github.com/xush6528

torch/fx/passes/annotate_getitem_nodes.py

@@ -23,7 +23,9 @@ def annotate_getitem_nodes(graph: torch.fx.Graph) -> None:
             # container types
             if hasattr(sequence_node.type, "_name"):
                 parameterized_types = sequence_node.type.__args__
-                if sequence_node.type._name == "Tuple":
+                if sequence_node.type._name == "Tensor":
+                    continue
+                elif sequence_node.type._name == "Tuple":
                     if len(parameterized_types) == 2 and isinstance(
                         parameterized_types[1], type(...)
                     ):