perf: improve bytes deserialization performance 2~3x

[v0y4g3r]

Changes

Specifically handle input than fits into one cache line, find out last non-continuation indicator and calculate the deserialized data size, we can also avoid reallocations.

Also add some roundtrip tests to assert correctness.

$ cargo bench  --bench 'serde' -- 'read_bytes' 
     Running benches/serde.rs (target/release/deps/serde-358374367734b2ff)
read_bytes/size-10      time:   [17.310 ns 17.408 ns 17.526 ns]
                        change: [-47.367% -47.052% -46.716%] (p = 0.00 < 0.05)
                        Performance has improved.

read_bytes/size-10-reverse
                        time:   [18.816 ns 19.252 ns 19.801 ns]
                        change: [-43.694% -41.106% -38.353%] (p = 0.00 < 0.05)
                        Performance has improved.

read_bytes/size-20      time:   [19.293 ns 19.390 ns 19.494 ns]
                        change: [-67.566% -67.285% -67.014%] (p = 0.00 < 0.05)
                        Performance has improved.

read_bytes/size-20-reverse
                        time:   [21.318 ns 21.457 ns 21.630 ns]
                        change: [-63.823% -63.175% -62.320%] (p = 0.00 < 0.05)
                        Performance has improved.

read_bytes/size-50      time:   [35.937 ns 36.147 ns 36.402 ns]
                        change: [-57.527% -57.064% -56.471%] (p = 0.00 < 0.05)
                        Performance has improved.

read_bytes/size-50-reverse
                        time:   [36.077 ns 36.240 ns 36.426 ns]
                        change: [-66.473% -66.063% -65.710%] (p = 0.00 < 0.05)
                        Performance has improved.