Add serde feature

[tidely]

Closes #23

Sounds simple enough! Well I'm here to make it complicated :)

Long story short, the default serde implementation for deserializing Arc<T> first allocates Box<T> followed by converting it to Arc<T> (which allocates again), this allows a blanket implementation for all ?Sized types which support deserializing into Box<T>. For Arc<str> for example, this allocates twice.

I've written our own Deserialize implementation, which now is specialized for Arc<str> and Arc<[T]>, while not supporting other !Sized types. I feel like this tradeoff is worthwhile, as we want to squeeze every last drop of performance out.

Tradeoff: Blanket Deserialize implementation which allocates twice for ?Sized vs optimized version which only works for implemented types, but only allocates once for ArcCow<'_, str>

I'd like some feedback on this!

Additionally I feel like since the serde feature adds a lot of code which is more complex than the rest of the crate, could it be time for some tests?