Optimize SocketMixin#write to avoid string copying

[casperisfine]

Ref: #962 (comment)

The key thing here is that byteslice is smart enough to create shared strings:

>> puts ObjectSpace.dump(PAYLOAD.byteslice(4096..-1).byteslice(4096..-1).byteslice(4096..-1))
{"address":"0x7fd1c93e0178", "type":"STRING", "class":"0x7fd1c90b7578", "shared":true, "encoding":"UTF-8", "references":["0x7fd1c93a3778"], "memsize":40, "flags":{"wb_protected":true}}

In such case only a RString is instantiated, the content isn't copied, so it's way faster. This was benched as 75x faster, however it's a synthetic benchmark without IOs.

cc @SamSaffron maybe you'd like to take a look?

[SamSaffron]

This is a fantastic change, I very much support it!

It reduces a bunch of code duplication and makes stuff much faster for large payloads which is awesome.

When I looked at this yesterday it "seemed" like stuff was structured in a way that we allowed people to override the _write_to_socket method, I think this is probably a bit too much flexibility especially it kind of follows from the naming that this is a private thing, but it may break some people so it is something to watch for.

That said, I 100% support this change and think we should go ahead with it!!!