[fx] Do not add Proxy on Tensor (pytorch#133470)

Summary: Switch to set_proxy_slot instead of set the proxy directly on the Tensor. We do not want to add Proxy to tensor objects, because Proxy cannot be deepcopied or pickeled and can cause problems when users want to deepcopy or pickle models.

Test Plan: CI

Differential Revision: D61277650

Pull Request resolved: pytorch#133470
Approved by: https://github.com/zou3519

torch/fx/experimental/proxy_tensor.py

@@ -886,23 +886,6 @@ def can_handle_tensor(x: Tensor) -> bool:
         name=proxy_mode.tracer.graph._target_to_str(func.overloadpacket.__name__),
     )
 
-    # This makes DCE marginally less likely to DCE inplace operations.
-    # It is not strictly necessary
-    # Kind of a hacky way to test if an op is in-place or not
-    if (
-        func.overloadpacket.__name__[-1] == "_"
-        and func.overloadpacket.__name__[0] != "_"
-    ):
-        if isinstance(args[0], List):
-            # e.g., c10d::allreduce_ returns a list of tensors as the first element
-            # in the output.
-            for i, a in enumerate(args[0]):
-                a.proxy = proxy_out[0][i]
-        else:
-            assert isinstance(args[0], Tensor), type(args[0])
-            # Adding an undefined attribute to Tensor?
-            args[0].proxy = proxy_out  # type: ignore[attr-defined]
-
     with _enable_thunkify(proxy_mode.tracer):
         out = func(*args, **kwargs)