[chore] use rayon work-stealing to improve evaluate_h

[jonathanpwang]

I have no idea how rayon work-stealing works, but by adding some par_iters to the multiple FFTs, it has improved the performance substantially on machines with large numbers of cores.

We have also discovered that Scroll's parallel FFT is better than the recursive FFT on c7a machines (8xl up to 48xl). So we use cfg(target_arch = "x86_64") to select which one to use now.