We already do this everywhere anyways, so this is more a correction
of the documentation than a change to the protocol. Therefore, I've
omitted describing any backwards compatibility handling.

Most messages have structured content (JSON, …) and require a lot more resources
to be parsed than is needed to only hold the byte buffer in memory.
Therefore, 4GiB is totally unrealistic and there is need for an
artificially reduced upper bound.

We use WebSockets as an example, but actually this applies to all protocols
that offer their own message framing

`relay-v2` is the same as `relay-v1`, but with enhanced hints encoding
that allows for different ways to connect to a relay server become
possible. Furthermore, relay hints may provide a human readable name.

Additionally, a few things are moved around between files, both
to put things where they belong and in anticipation of future
features.