Experiment: Configurable Pickle protocol version

[moreati]

Mitogen currentl uses Pickle protocol 2, regardless of the Python versions. This constrains performance under Python 3 (#485). Configurable Pickle protocol would improve performance when exclusively using Python 3

Protocol 2

➜  mitogen git:(mitogen.pickle_protocol) ✗ .venv/bin/python3.14 tests/bench/throughput.py --ssh-host d13.lan --ssh-python python3 
2
Testing local()...
local() took 2734.70 ms to transfer 512.00 MiB, 187.22 MiB/s
Testing ssh(compression=False)...
ssh(compression=False) took 36714.61 ms to transfer 512.00 MiB, 13.95 MiB/s
Testing ssh(compression=True)...
ssh(compression=True) took 43327.35 ms to transfer 512.00 MiB, 11.82 MiB/s

Protocol 3

➜  mitogen git:(master) ✗ MITOGEN_PICKLE_PROTOCOL=3 .venv/bin/python3.14 tests/bench/throughput.py --ssh-host d13.lan --ssh-python python3
3
Testing local()...
local() took 324.69 ms to transfer 512.00 MiB, 1576.91 MiB/s
Testing ssh(compression=False)...
ssh(compression=False) took 22553.84 ms to transfer 512.00 MiB, 22.70 MiB/s
Testing ssh(compression=True)...
ssh(compression=True) took 21976.54 ms to transfer 512.00 MiB, 23.30 MiB/s

[moreati]

SSH benchmarks were over WiFi

[moreati]

Cousin of #1416, knowing python version of contexts would allow dynamically choosing best supported Pickle protocol.

[moreati]

Little to no speed-up of single-shot benchmark

➜  mitogen git:(mitogen.pickle_protocol) ✗ hyperfine --warmup 1 'python3.14 hello_local2.py' 'PICKLE_PROTOCOL=3 python3.14 hello_local2.py'        
Benchmark 1: python3.14 hello_local2.py
  Time (mean ± σ):     190.0 ms ±   1.9 ms    [User: 119.3 ms, System: 20.9 ms]
  Range (min … max):   187.1 ms … 194.4 ms    15 runs
 
Benchmark 2: PICKLE_PROTOCOL=3 python3.14 hello_local2.py
  Time (mean ± σ):     188.7 ms ±   1.8 ms    [User: 118.3 ms, System: 20.6 ms]
  Range (min … max):   186.4 ms … 192.2 ms    15 runs
 
Summary
  PICKLE_PROTOCOL=3 python3.14 hello_local2.py ran
    1.01 ± 0.01 times faster than python3.14 hello_local2.py