Change check to use torch.compiler.is_compiling()

Summary:
Fixes D82792378, where we were creating an op in the graph while unflattening the jaggedtensor
I think we should just change all the is_torchdynamo_compiling to torch.compiler.is_compiling since they are gating similar behavior/assumptions in PT2

Reviewed By: TroyGarden

Differential Revision: D82837368

fbshipit-source-id: 63fa1aff39207b8e3dc1d548b9165f595bfccf65

Author: angelayi

torchrec/sparse/jagged_tensor.py

@@ -1066,7 +1066,7 @@ def _jt_flatten_spec(t: JaggedTensor, spec: TreeSpec) -> List[Optional[torch.Ten
 def _assert_tensor_has_no_elements_or_has_integers(
     tensor: Optional[torch.Tensor], tensor_name: str
 ) -> None:
-    if is_torchdynamo_compiling() or tensor is None:
+    if torch.compiler.is_compiling() or tensor is None:
         # Skipping the check tensor.numel() == 0 to not guard on pt2 symbolic shapes.
         # TODO(ivankobzarev): Use guard_size_oblivious to pass tensor.numel() == 0 once it is torch scriptable.
         return