[pytorch] Disable fast path in MultiheadAttention in Export (pytorch#106824)

Summary:
We are seeing `aten._native_multi_head_attention` op (not in core Aten op set) is left in the exported graph and causes problems in the downstream at runtime.

Two proposed solutions:
 1. Disable fast path while tracing to leverage the non-optimized path to get decomp, that way, the blamed op won't show up in the exported graph
 2. Add a decomp rule for `aten._native_multi_head_attention`

After discussing with kimishpatel and bdhirsh, #1 is preferred and verified it could immediately unblock the critical model enablement work for PP.

Test Plan: CI

Differential Revision: D48169806

Pull Request resolved: pytorch#106824
Approved by: https://github.com/kimishpatel

Author: guangy10

torch/nn/modules/activation.py

@@ -895,6 +895,14 @@ def _arg_requires_grad(x: Optional[torch.Tensor]) -> bool:
     return False
 
 
+def _is_make_fx_tracing():
+    if not torch.jit.is_scripting():
+        torch_dispatch_mode_stack = torch.utils._python_dispatch._get_current_dispatch_mode_stack()
+        return any(type(x) == torch.fx.experimental.proxy_tensor.ProxyTorchDispatchMode for x in torch_dispatch_mode_stack)
+    else:
+        return False
+
+
 class MultiheadAttention(Module):
     r"""Allows the model to jointly attend to information
     from different representation subspaces as described in the paper:
@@ -1169,6 +1177,8 @@ def forward(
             # generator expressions.
             if torch.overrides.has_torch_function(tensor_args):
                 why_not_fast_path = "some Tensor argument has_torch_function"
+            elif _is_make_fx_tracing():
+                why_not_fast_path = "we are running make_fx tracing"
             elif not all(_check_arg_device(x) for x in tensor_args):
                 why_not_fast_path = ("some Tensor argument's device is neither one of "
                                      f"cpu, cuda or {torch.utils.backend_registration._privateuse1_backend_name}")