[dynamo][compile-time] Remove preserve rng state per op (#121923)
We already have one globally - https://github.com/pytorch/pytorch/blob/02bb2180f497d6d10dd11da51fffefaa5af80aca/torch/_dynamo/convert_frame.py#L477
I don't think we need per op.
Saves ~2 seconds on this benchmark
~~~
def fn(x):
for _ in range(10000):
x = torch.ops.aten.sin(x)
return x
~~~
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121923
Approved by: https://github.com/jansel