fwd_rng_state show up in the aot_export_joint grpah input

[SherlockNoMad]

See pytorch/torchtitan#1794

P1975157784: rank0_autograd_function_0fea2786.py

Setting torch._functorch.config.graphsafe_rng_functionalization = False doesn't work.

How to avoid fwd_rng_state from showing up?

cc @chauhang @penguinwu @zou3519 @bdhirsh