Several performance improvements to serialization

[SabrinaJewson]

• The itoa and ryu crates (the same crates serde_json uses) avoid allocating anything when serializing integers and floats.
• By switching the code to work on bytes instead of &strs, expensive Unicode handling can be avoided (e.g. str::bytes is a lot cheaper to run than str::chars).
• We can copy serde_json and start serialization with a capacity of 128, to avoid excessive reallocations.
• Since we use bytes now we might as well expose ser::to_bytes.
• Edit: I came back to this and overrode Serializer::collect_str because we can provide a more efficient implementation than the default.