Use const generics in Dirichlet

[Armavica]

This is a draft PR for issue #1006.

It is a working Dirichlet distribution that takes and returns const arrays, working with rand's current MSRV (1.56) [edit: but apparently not with Serde: serde-rs/serde/issues/1937].

It might still be useful to keep the original dynamic Dirichlet, so should I leave alone the original Dirichlet distribution and create a new one such as DirichletConst?

[dhardy]

Seems like a good change.

Add a note to the CHANGELOG please.

[Armavica]

serde won't serialize const arrays by default but I found a way to do it anyways with the crate serde_arrays: there might be other solutions which I would be happy to try if you want, let me know what you think!

[dhardy]

The crate doesn't appear that well maintained. I'm guessing testing rand_distr with --no-default-features --features serde1 will fail now?

Perhaps an alternative is to serialise the array as a slice?

[Armavica]

The crate doesn't appear that well maintained. I'm guessing testing rand_distr with --no-default-features --features serde1 will fail now?

No, in the current state of the PR, this testing configuration succeeds.

Perhaps an alternative is to serialise the array as a slice?

Sure, I can try that.

[Armavica]

The crate doesn't appear that well maintained.

I found another crate which looks much more active: serde_with (PR updated).

I think that serializing as a slice would require to write our own serializing/deserializing code, right?

[dhardy]

This looks okay, but I suggest serde1 should not imply serde_with. Instead, use serde_with as the feature-gate for serialization on Dirichlet.

[Armavica]

Is this what you meant? I wasn't sure how the name the feature. I first tried to go with serde_with but apparently one cannot have a feature with the same name as the dependency. Or perhaps the feature name should mention Dirichlet or arrays?

[dhardy]

but apparently one cannot have a feature with the same name as the dependency

You can. The old way to do this is just to have the optional feature in the Cargo.toml (you can still use feature = "serde_with" in code). The new way (since 1.60) is namespaced features, e.g. serde_with = ["dep:serde_with"] in the Cargo.toml. Our MSRV is still 1.56 so we have to use the former.

[vks]

Looks great, thanks!