Consider adding next_f32/next_f64 methods to rand::Rng

[huonw]

There are random number generators that generate floating point numbers directly (e.g. dSFMT). Currently, implementing the rand::Rng trait requires next_u32 and possibly next_u64 (which must be implemented to be uniformly distributed) so these random number generators require some very inefficient contortions to fit into this interface, and don't get any speed benefit when generating f32s or f64s since everything goes via next_u32 and next_u64.

I propose changing rand to:

impl Rand for f32 {
    fn rand<R: Rng>(rng: &mut R) -> f32 {
        rng.next_f32()
    }
}

trait Rng {
    // ...

    /// Override if your RNG can generate [0, 1) numbers more efficiently.
    fn next_f32(&mut self) -> f32 {
        (rng.next_u32() >> 8) as f32 / (1 << 24) as f32
    }
}

and similarly for f64. This means a RNG that creates floating point numbers directly can override next_f32 and next_f64 so that generic users reap the benefits.

Random numbers are often used for numeric simulations, and so this seems like a sensible addition, and I cannot think of any other core types that deserve an explicit specialisation in this form so this shouldn't be a slippery slope.