benchmarking suite should initialize cuda graphs / profiler interaction

[tfogal]

As per: https://pytorch.org/docs/stable/torch.compiler_profiling_torch_compile.html#working-around-cuda-graph-profiling-issues, we may need to do some initialization when using cuda graphs.

We are not yet using cuda graphs, but the benchmarking code should just invoke this at the start of execution anyway. Thus if we add a benchmark that graphs around something nvFuser gives, or if we start internally using graphs down the road, we won't hit surprising profiling issues.