Better support for multivariate distributions

[vks]

The Distribution trait does not work well for multivariate distributions, because impl Distribution<Vec<f64>> for Multivariate requires an allocation for every sample (see #485 for an example). Maybe we should add another trait:

pub trait MultivariateDistribution<T> {                                                         
    fn sample_multi<R: Rng + ?Sized>(&self, rng: &mut R, buffer: &mut [T]);
}

With SIMD in mind, this might even make sense for univariate distribution. So we might want to add sample_multi to the Distribution trait.

[dhardy]

This is a good point. My preference is to postpone this until we resolve #290 or have more multi-variate distributions.

As mentioned in #485 there might be another solution, if fixing the number of variables at compile-time is sufficient: using generic consts (rust-lang/rust#44580).

[vks]

Nice, I did not know const generics were so close!

rust-lang/rust#51192 (comment)

Given that const generics on nightly might be only a month away, my vote would be to wait...

[dhardy]

We now use const generics: impl<F, const N: usize> Distribution<[F; N]> for Dirichlet<F, N> where ...

Caveat: we currently use Vec to construct some arrays in distribution constructors. Once core::array::try_from_fn is stable, we can use that instead.

Caveat: sometimes arrays of other lengths are required (DirichletFromBeta uses an internal array of length one less than the output length). This isn't really an issue, but currently uses an allocator (doing otherwise would require generic_const_exprs or an empty element).

Caveat: the above approach does not support run-time variable length N. We likely don't need to care about this.

---

Conclusion: we can improve our code once try_from_fn is stabilized, but probably not beyond that.

I tried simply copying the code behind try_from_fn as a utility fn; this is viable but requires rather more unsafe code than I'd like (besides which, we forbid unsafe from rand_distr).

[dhardy]

Closing: #1006 tracks the remaining issues in Dirichlet, and I don't think there's anything else we want for Distribution.