Use zstd for multiprocess IPC compression when available

[akx]

A lot of time was being spent in zlib_compress; on modern Pythons, we can use zstd for free.

Refs #674 (and conflicts with it).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (db53893) to head (b95b396).

Additional details and impacted files

@@             Coverage Diff             @@
##           master      #676      +/-   ##
===========================================
+ Coverage   99.95%   100.00%   +0.04%     
===========================================
  Files          20        21       +1     
  Lines        2473      2488      +15     
===========================================
+ Hits         2472      2488      +16     
+ Misses          1         0       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[gaogaotiantian]

Do you have benchmark to support the idea?

[akx]

@gaogaotiantian Sure. Looks to be 4% faster on a trivial small example, likely more on a real program:

$ hyperfine 'git checkout master && .venv/bin/viztracer foot.py' 'git checkout zstd-when-possible && .venv/bin/viztracer foot.py'
Benchmark 1: git checkout master && .venv/bin/viztracer foot.py
  Time (mean ± σ):      2.707 s ±  0.034 s    [User: 2.495 s, System: 0.201 s]
  Range (min … max):    2.664 s …  2.756 s    10 runs

Benchmark 2: git checkout zstd-when-possible && .venv/bin/viztracer foot.py
  Time (mean ± σ):      2.603 s ±  0.022 s    [User: 2.393 s, System: 0.201 s]
  Range (min … max):    2.565 s …  2.643 s    10 runs

Summary
  git checkout zstd-when-possible && .venv/bin/viztracer foot.py ran
    1.04 ± 0.02 times faster than git checkout master && .venv/bin/viztracer foot.py