Turn off cudagraph trees (#98709)
There were some recent failures on master, and I think it's fair to defer on turning it on till we get a bit of the Tensor construction overhead down because that shows up a lot in the TB benchmarks.
There may ultimately be an unavoidable tradeoff between memory and performance to some extent but we can get the overhead numbers down a bit first.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98709
Approved by: https://github.com/Chillee