feat(ed255519): Introduce serde_bytes optional dependency

[connec]

This allows consumers to opt-in to serde_bytes support by setting the
std, serde, and serde_bytes features. This makes it convenient to
optimise (de)serialization for formats with efficient byte array
encodings, e.g.

#[derive(serde::Deserialize, serde::Serialize)]
struct MySignature {
    #[serde(with = "serde_bytes")]
    signature: ed25519::Signature,
}

Without first-class support for serde_bytes, consumers would have to
write their own (de)serialization machinery for this common case.

[connec]

In contrast to #338 this has the advantage of keeping the features orthogonal, so --all-features still works fine and testing is kept simple. The downside is an additional dependency in the tree, and more complex integration for consumers (e.g. the need for an extra attr).

[tarcieri]

@connec I think overall I prefer this approach to one which introduces crate features which don't properly unify i.e. #338

It does introduce an extra dependency, but I'm okay with that given the provenance of that particular dependency (i.e. it's by serde's authors and therefore is as trustworthy as serde itself)

[connec]

I've pushed an update that removes the compile_error! and the dependence on TryFrom<Vec<T>> for [T; N]. For the latter, based on the docs I've just removed the visit_byte_buf method, since there's no way to create a >32 element array from a Vec with a copy as far as I can tell.

[tarcieri]

Looks good, thank you!

It might be good to add the example from the description to the rustdoc as well.

[tarcieri]

Released in #340