hybrid-array: Added serde impls for Array

[rozbb]

I'd love to switch hpke over to hybrid-array, since it's becoming pretty hard to manually audit generic-array, and we expose it so it's not like we can get away with leaving it old forever.

One of the really convenient things that generic-array provides, though, is Serialize/Deserialize impls for GenericArray. We use this in our Xyber impl. The only other option for autoderiving these traits is serde-bigarray (upstream support is not coming any time soon), and I'd rather not bring it in.

This PR should bring hybrid-array a little closer to feature-parity with generic-array. I don't love the unsafe stuff here, but it's pretty minimal and well-documented, and also has a clear roadmap for removing it once certain functions stabilize.

[rozbb]

It looks like some tests whose MSRV hasn't been bumped have a problem with this Cargo.toml

[tarcieri]

We've been there and back again with the various tradeoffs of serde support across formats. See the recently merged:

RustCrypto/formats#1111

RustCrypto/formats#1112

There is so much errata involved here across various serde-based format implementations that I have trouble keeping it all in my head. The array serializers look like this: https://github.com/RustCrypto/formats/blob/master/serdect/src/array.rs

We tried to use the tuple serializers initially to avoid the length prefix, but it turns out those have terrible interactions with MessagePack because serde lacks the ability to describe homogenously typed tuples. serialize_bytes exists specifically for handling the byte array case, which is far and away our biggest use case, but there is no general-case solution for fixed size arrays: serde-rs/serde#2120 (comment)

I guess the question is whether we should worry about these particular issues in the context of this crate, namely formats like MessagePack will do a lot of branching on the data and encode it oddly because they have to encode type information for each entry in the tuple, rather than storing it in one place.

p.s. I am very, very tired of reviewing serde-related code. Just when I think I understand what's happening I discover yet another fun new surprise.

[tarcieri]

We can definitely bump MSRV on fiat-constify, FWIW.

[rozbb]

I'm kind of in agreement here. I think we don't necessarily have to expect [u8; N] to be serializable. Or at the very least I can live without it and just write my own impls for my use case.

If you want to keep ArrayExt::try_from_fn, then I can remove the rest of this PR. Otherwise, feel free to close outright.

[rozbb]

In this same vein. I thing maybe hpke shouldn't be doing serde impls either, for the same reason. It has to_bytes and from_bytes. I think that should be sufficient. If someone wants a specific format then they can choose.

[tarcieri]

hybrid-array has moved to https://github.com/RustCrypto/hybrid-array