Massive save state speedup

[Dr-Emann]

Because libretro has an API that asks for the size seperately, and we
have to serialize to get the size, the end result is that we change from
2 serializations for each save state down to 1.

[Dr-Emann]

I found a huge speedup: the overhead of serializing all the u8's individually in the the large Box<[u8]>s was huge. Serde has a solution to this: serializers have a serialize_bytes function (and bincode implements it very efficiently), however, until specialization lands, serde can't call it automatically. So just annotating the Box<[u8]> fields with #[serde(with = "serde_bytes")] causes the fields to be serialized with a single serialize_bytes call, instead of a ton of individual calls per u8.

On my system, this took the time to serialize from over 100ms to about 5ms.

[yupferris]

Thanks for the PR! I like these changes :)

A couple things I'd like changed, but overall great work; once the nits have been addressed I'd love to go over this and merge it on the next stream :)

[Dr-Emann]

I believe I've addressed the nits. It's pretty cool to be able to turn on rewind, and not bring things to a grinding halt.

[yupferris]

Fantastic, thanks! I'll go over this and merge on the stream this evening. :)