Rework as Bytes-only payload

[alexheretic]

Spurred by the chat in tokio-rs/axum#3078 This PR experiments with simplifying the API to use only Bytes as payload. It does work pretty well, the only downside seems to be client->server write performance as masking requires mutating the bytes. Now fixed.

• Read unmask performance was addressed by unmasking before converting to Bytes to avoid the overhead of having to mutate them.
• Write masking performance was addressed by writing and unmasking in the out_buffer rather than in-place.

[alexheretic]

I wonder if there is a way of optimising the write to not need to convert Bytes too...

[agalakhov]

Performance regression at 100k small messages seems to be significant. This however does not mean that it will be so in real use. I would test it a little bit more in other environments. The real question is, how many Arcs exist in real life and how many of them do exist only due to Bytes.

Such a performance impact (10%) in a real-life high load example would be huge. Ten percent more load means ten percent more servers in the cluster and ten percent more hardware costs, and hardware is expensive.

[alexheretic]

Actually I found we don't need to apply the mask then write, we have an out_buffer so we can write into the out_buffer then apply the mask to the written slice. This has no extra cost and means we have no need to mutate the payloads at any point.

[alexheretic]

Write performance looks good now

write 100k small messages then flush (server)
                        time:   [4.7412 ms 4.7432 ms 4.7437 ms]
                        change: [+2.1394% +2.4145% +2.6909%] (p = 0.08 > 0.05)

write+mask 100k small messages then flush (client)
                        time:   [5.9248 ms 5.9282 ms 5.9416 ms]
                        change: [-1.9907% -1.5423% -1.0910%] (p = 0.05 < 0.05)