feat: 35% faster decompression with less boundary check

[XiangpengHao]

Hi SpiralDB devs! First of all, thank you for implementing the great FSST library!

While profiling a decompression workload, I noticed we have some unintended boundary checks that could be removed to improve decompression throughput by 35%, as shown in the decompress benchmark below.

Removing those boundary checks requires more unsafe code, I've mannually annotated their correctness, but please double check! cc @alamb who might be interested

cargo bench --bench micro decompress

   Compiling fsst-rs v0.4.2 (/home/hao/coding/fsst)
    Finished `bench` profile [optimized] target(s) in 0.86s
     Running benches/micro.rs (target/release/deps/micro-bf2c70cedbbe0467)
Gnuplot not found, using plotters backend
cf=8/decompress         time:   [47.687 µs 47.725 µs 47.768 µs]
                        thrpt:  [20.444 GiB/s 20.462 GiB/s 20.478 GiB/s]
                 change:
                        time:   [-26.409% -26.311% -26.208%] (p = 0.00 < 0.05)
                        thrpt:  [+35.517% +35.705% +35.887%]
                        Performance has improved.
Found 11 outliers among 100 measurements (11.00%)
  6 (6.00%) high mild
  5 (5.00%) high severe

[CLAassistant]

All committers have signed the CLA.

[alamb]

FYI @a10y

[a10y]

Thanks @XiangpengHao ! Just two small comments (perhaps I'm being overly paranoid about add vs byte_add but just prefer to be explicit).

[XiangpengHao]

Thanks for the review! I've updated the code and rerun the benchmark and see no perf difference (just to be sure).

[a10y]

@XiangpengHao this change is now published in 0.4.3 https://crates.io/crates/fsst-rs/0.4.3