gh-144764: Made dataclasses construct automatic docstrings lazily

[adamchainz]

Move the default docstring construction to occur on-access through a descriptor.

Verified with tprof and this script that generates 10k dataclasses:

from dataclasses import dataclass

for i in range(10_000):
    @dataclass
    class Example:
        field1: int
        field2: str
        field3: float

Before:

$ tprof -t dataclasses._process_class example.py
🎯 tprof results:
 function                     calls total  mean ± σ       min … max
 dataclasses._process_class() 10000    5s 485μs ± 120μs 458μs … 6ms

After:

$ PYTHONPATH=Lib/ uvx tprof -t dataclasses._process_class example.py 🎯 tprof results:
 function                     calls total  mean ± σ       min … max
 dataclasses._process_class() 10000    3s 275μs ± 131μs 245μs … 6ms

The mean time spent in _process_class() has dropped from 485μs to 275μs, a ~42% time saving (admittedly skewed due to the small size of the dataclass).

• Issue: Make dataclasses construct automatic docstrings lazily #144764

[johnslavik]

DocDescriptor isn't meant to be public API. Can we rename it to a private class?

[adamchainz]

DocDescriptor isn't meant to be public API. Can we rename it to a private class?

Good point, done!

[danielhollas]

Hehe, funny, I ended doing do same trick in my recent PR #144387, with the main purpose of being able to lazy import the inspect module. It's great to see it improves data class creation as well (I had a hunch but didn't do careful benchmarking).