Ship stable-ABI (abi3-py311) wheels — one wheel per platform, every CPython ≥ 3.11

[kurok]

Summary

Implements the abi3 half of #101: wheels now target the CPython stable ABI (cp311-abi3), so one wheel per platform covers every CPython ≥ 3.11 — including versions that don't exist yet. This permanently ends the "new Python, no wheel" failure mode behind #14 and #15.

Changes

• Cargo.toml — enable pyo3/abi3-py311 (baseline matches requires-python >= 3.11).
• publish.yml — drop --find-interpreter: each platform job builds a single abi3 wheel instead of one per interpreter. Platform matrix unchanged.
• test.yml — build once, test everywhere: a new build-wheel job builds one manylinux abi3 wheel; the 3.11–3.14 test matrix installs that same wheel and runs the suite against it. The abi3 promise is verified, not assumed. Side effect: the test matrix no longer needs a Rust toolchain — four fewer Rust builds per PR.
• Readme.md — supported-versions table + free-threaded note.
• CHANGELOG.md — Unreleased entry.

Benchmark decision (the >5% rule from #101)

Min parse time, CPython 3.12, Apple Silicon (M-series), two runs per configuration, --benchmark-min-rounds=25:

Build	run 1	run 2	speedup vs mail-parser
version-specific (master)	2.0939 ms	2.1197 ms	4.36x / 4.32x
abi3 (this PR)	1.8786 ms	1.8474 ms	4.92x / 5.01x

abi3 measured ~11% faster, consistently across runs. Decision: full abi3, no hybrid version-specific wheels. The check_benchmark.py floor (7.0x) is left unchanged: the CI benchmark job builds from source, which is now abi3 by default, and the measured direction is favorable.

Free-threaded CPython (deferred, with reasons)

The cp313t/cp314t half of #101 is deferred:

• Free-threaded CPython does not support the stable ABI, so ft wheels are a separate version-specific build axis — orthogonal to this change.
• Running without the GIL needs an explicit PyO3 declaration (gil_used = false) plus a shared-state audit. Per abi3 stable-ABI wheels + free-threaded CPython support (end the new-Python-no-wheel cycle) #101 that audit must coordinate with the planned lazy-parsing work (Lazy parsing modes: metadata-only sweeps and deferred attachment decoding (mode="metadata"|"lazy") #97), which introduces the main shared-state risk — auditing before it lands would have to be redone.
• Current behavior on 3.13t/3.14t (source build, GIL re-enabled on import) is documented in the README.

Verification

• One locally built cp311-abi3-macosx_11_0_arm64 wheel installed on CPython 3.11, 3.12, and 3.14: 91/91 tests pass on each (3.13 is covered by the CI matrix).
• actionlint and YAML parse clean on both workflows; cargo fmt --check clean.

Cross-refs #14 and #15 — those close when the fixed release actually reaches PyPI (blocked on the PEP 541 transfer, pypi/support#11044).