Make modint::Barrett (AtomicU32, AtomicU64)

[qryxip]

Benchmark:
https://gist.github.com/qryxip/c17bdb69126170900ed4333f0bf3198d

[koba-e964]

I'm not sure this PR is implemented as intended, but getting nearly 2x faster by AtomicXX is something I never dreamed of. Great!

[qryxip]

In fact I didn't intend performance improvement. Or rather I wanted to know how much it decreases. I was thinking that we should write some document to prevent users from doing this:

use std::{panic, thread};

ModInt::set_modulus(m);

thread::Builder::new()
    .stack_size(16 * 1024 * 1024)
    .spawn(|| {
        todo!();
    })
    .unwrap()
    .join()
    .unwrap_or_else(panic::resume_unwind);

[qryxip]

I've just found there's almost no overhead compared with this:

pub struct Barrett(UnsafeCell<BarrettInner>);

struct BarrettInner {
    m: u32,
    im: u64,
}

/// # Safety
///
/// Nothing. Do not ever do this.
unsafe impl Sync for Barrett {}