SkewNormal distribution implementation

[saona-raimundo]

Related to #1149.

Implements the Skew Normal distribution.

Most notably I had to add a dev-dependency to include distribution tests similar to #1121.
Adding the test was suggested by @dhardy.
There was a brief discussion about adding dev-dependencies for special functions with @vks in here. We talked about the crate spfunc, since it is a pure Rust implementation of special functions. It is still a very young crate and the function erf is still not there. Therefore, I chose to add statrs as it is a close companion of this crate, which is also pure Rust.

[saona-raimundo]

@dhardy, can I ask for help with these tests?
Could it be a problem with the quote crate and its use of constant generics?

[vks]

The nightly failures are unrelated. The other failure seems to be in nalgebra, which I think is pulled in by statrs.

[vks]

I wonder whether special is a better choice than statrs in this case. It also provides the error function, but has less dependencies.

How does the sampling work for infinite location? It looks like it would just always return the infinite value.

[vks]

We should probably figure out how to deduplicate the code here, but this can be done in a follow-up PR.

[saona-raimundo]

Yes! Although I am not so sure how to do the code fully generic to test discrete distributions in the future.

For now, for these continuous distributions, I would put together the repetitive code in a function that takes the variables n_samples, min_x, max_x, dist, pdf and name (name of the distribution being tested for reporting it).

It would look like this:

#[test]
fn skew_normal() {
    const N_SAMPLES: u64 = 1_000_000;
    const LOCATION: f64 = 2.;
    const SCALE: f64 = 0.5;
    const SHAPE: f64 = -3.0;
    const MIN_X: f64 = -1.;
    const MAX_X: f64 = 4.;

    let dist = SkewNormal::new(LOCATION, SCALE, SHAPE).unwrap();
    fn pdf(x: f64) -> f64 { 
        // some code
    }
    histogram_test(N_SAMPLES, MIN_X, MAX_X, dist, pdf, "skew_normal");
}

where histogram_test takes the code repeated in both cases.

You see that the signature of histogram_test is not so clear for dist when testing discrete variables. And one would have to take into account the width of the bins of the histogram in that case too.

[saona-raimundo]

How does the sampling work for infinite location? It looks like it would just always return the infinite value.

Yes, indeed! It will return infinite (which is consistent with Normal and the mathematical definition). I added tests for this case too.

[saona-raimundo]

I wonder whether special is a better choice than statrs in this case. It also provides the error function, but has less dependencies.

I changed to special, but the Maybe nightly tests still fail.

[saona-raimundo]

By the way, @vks, thank you for the review!! :D

[vks]

This looks good!

The nightly failures are unrelated (packed_simd_2 is broken on the current nightly, this happens from time to time).

I think this can be merged.