Attempt to split generated u64 values into two u32

This *reduces* performance in the mt19937_64::benchmark_gen_624_u32 benchmark
in my tests.

Pushing for comment, DO NOT MERGE.

src/mt19937_64.rs

@@ -27,12 +27,14 @@ const LM: Wrapping<u64> =  Wrapping(        0x7FFFFFFF); // Least significant 31
#[allow(non_camel_case_types)]
#[derive(Copy)]
pub struct MT19937_64 {
    next_u32: Option<u32>,
    idx: usize,
    state: [Wrapping<u64>; NN],
}

const UNINITIALIZED: MT19937_64 = MT19937_64 {
    idx: 0,
    next_u32: None,
    state: [Wrapping(0); NN]
};

@@ -92,7 +94,13 @@ impl<'a> SeedableRng<&'a [u64]> for MT19937_64 {
impl Rng for MT19937_64 {
    #[inline]
    fn next_u32(&mut self) -> u32 {
        self.next_u64() as u32
        if let Some(n) = self.next_u32.take() {
            n
        } else {
            let n64 = self.next_u64();
            self.next_u32 = Some((n64 >> 32) as u32);
            n64 as u32
        }
    }

    #[inline]