Enable aten-aten decomps (#85921)
Invokes aten-aten decomps with re-entrant FakeMode. These decomps are being used in other places, so it's good to unify the path static fake tensor takes / get additional testing etc. There is also an instance where we return different devices with cpu/cuda which this fixes ([batch_norm](https://github.com/pytorch/pytorch/blob/master/torch/_decomp/decompositions.py#L1374))
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85921
Approved by: https://github.com/ezyang
