Support bytes::Bytes & bytes::BytesMut payloads for binary & text messaging

[alexheretic]

This PR builds on top of #462

• Adds Utf8Payload for text messages backed by Payload.
• Tri-variant Payload for Bytes, BytesMut & Vec.
  • Bytes gives flexibility supporting shared data & cheap cloning.
  • BytesMut provides best read performance as we can get these from the new read buffer sans copying & unmask optimially too.
  • Vec provides best write performance. I couldn't get BytesMut to perform as well even changing the bench user code. Since it doesn't increase Payload 40 bytes size it seems worth having.
• Replace the read buffer with a BytesMut based one that can then emit payloads without copying.
• Optimise incomplete message handling so complete messages do not need to be copied (general case for small messages).
• Add Payload::share for cheap cloning.
• Make Payload::eq compare data, so allow equality across variants.
• Remove and change a bunch of fns & types to suit the new internals.
• Add general &[u8]/&str From impl and use const fn from_static to construct static Payload/Utf8Payload data.

The main advantage here is extending the bytes flexibility to Text types and better read performance.

Benchmarks

Write performance remains a little slower, no more than noted in the parent PR. However, read performance is significantly faster.

$ cargo bench --bench \* -- --quick --noplot --baseline master
read+unmask 100k small messages (server)
                        time:   [7.8768 ms 7.9103 ms 7.9187 ms]
                        change: [-38.736% -37.861% -36.967%] (p = 0.06 > 0.05)

read 100k small messages (client)
                        time:   [7.1168 ms 7.1322 ms 7.1939 ms]
                        change: [-34.273% -33.840% -33.406%] (p = 0.07 > 0.05)

write 100k small messages then flush (server)
                        time:   [4.6548 ms 4.6852 ms 4.6928 ms]
                        change: [+6.0437% +6.5546% +7.0663%] (p = 0.10 > 0.05)

write+mask 100k small messages then flush (client)
                        time:   [6.0834 ms 6.0875 ms 6.1039 ms]
                        change: [+6.0147% +6.6040% +7.1980%] (p = 0.10 > 0.05)

[daniel-abramov]

You're fast! 🙂

Interesting decision to switch from ReadBuffer to BytesMut for the in_buffer! It's nice that it helps to spare additional allocations (like that one with a vector that had been previously allocated just to read and return a payload from the in_buffer).

[Zarathustra2]

Wow that is awesome!

@daniel-abramov Do you have an ETA when you plan on releasing 0.25 for tungstenite & tokio-tungstenite ?

[daniel-abramov]

Wow that is awesome!

@daniel-abramov Do you have an ETA when you plan on releasing 0.25 for tungstenite & tokio-tungstenite ?

Published! 🎉