If we update the state, or indicate a pessimistic fixpoint, we need to consider NestedParallelism too. Fixes part of https://github.com/llvm/llvm-project/issues/66708 That said, the reproducer still needs malloc which we don't support on AMD GPU. Will be added later.