build(labels): bump demangle to 0.6.0, take #383 review nits

[r0ny123]

Follow-up to #383, which merged before these were pushed.

Before: smda pins demangle==0.5.1. The pin comment does not say what a host application in the same process can still do, and the #379 CHANGELOG entry has no PR links, a vaguer module reference and a cost figure without its baseline.

After: smda pins demangle==0.6.0. The pin comment notes that a host calling load_plugins() or register_language() can still replace a built-in scheme, leaving Itanium, MSVC and Swift names mangled without an error. The CHANGELOG entry links (#379, #383), names the four modules, and states the cost against the native build, marked as measured on 0.5.0.

0.6.0 does change schemes/ (MSVC operator<=>, ARM64e qualifiers, Swift ~Deinitable, and the 0.5.2 bounds fixes for oversized Rust lengths and non-ASCII Swift identifiers). Every name the test suite passes to demangle, 20,430 of them, gives the same output and the same exceptions on 0.5.1 and 0.6.0. Entry points are still read only in load_plugins(), testDemangleSchemes passes, and there are no new dependencies.

Full suite: 2,244 passed, 2 skipped. ruff, ruff format and ty are clean.

[danielplohmann]

The bump itself is fine. I diffed the 0.5.1, 0.5.2 and 0.6.0 wheels. Entry points are still read only in load_plugins(). There are no new dependencies and no subprocess or network calls. On smda's inputs:

• Fixtures: 8,282 function labels from 46 analysed fixtures, 0 changed.
• Replay: all 22,321 distinct (scheme, name) pairs from two suite runs plus the fixtures give the same result on both versions.
• Fuzzing: about 8,950 mutated and pathological names, with no crash or timeout. The slowest name drops from 0.22 s to 0.033 s.

What needs changing is how the bump is described, because 0.6.0 does change labels, just not on our fixtures:

• Itanium and Swift (0.5.2): source-name lengths count UTF-8 bytes instead of characters. _ZN4main5caféE now reads as main::café where it was refused, _Z4cafév is now refused, and $s4main5caféV1fyyF reads as main.café.f() -> () instead of garbage. That's correct, since GCC/Clang mangle byte lengths, but non-ASCII C++/Swift names get new labels.
• MSVC ARM64e __ptrauth (0.6.0): new grammar; these names used to be refused.
• Swift ~Deinitable is now rendered by name.
• Rust: bounds checks and a faster path for oversized lengths, with no output difference found.
• MSVC operator<=>: no change. _DOUBLE_UNDERSCORE_OPERATORS is identical in 0.5.1 and 0.6.0, so please drop that from the description.

So:

1. CHANGELOG: please restore #379's bullet (pin aside) and give this bump its own (labels) bullet, under Changed or Compatibility since labels move on binaries with non-ASCII or ARM64e names. Name the changes above and the measurement: "8,282 fixture labels, 0 changed, reproduced by reviewer; no fixture name is non-ASCII". As written, #379's "every Itanium, MSVC and Rust label is unchanged" was measured on 0.5.0 and no longer holds. Please also drop the 20,430: Hypothesis draws new names every run (19,014 and 19,015 on two runs, 15,709 shared), so the count can't be reproduced. Then link (#387).
2. Pin comment (pyproject.toml): a replacing plugin that parses would silently change labels, not leave them mangled. Something like "can replace a built-in scheme and change or drop Itanium, MSVC and Swift labels without an error".

Upstream nits, not for this PR: __ptrauth data symbols render without a space before the name (int *__ptrauth(2, 3, 1)p), and ptrauthNumber accepts any number of nibbles and wraps at 64 bits.

With the CHANGELOG fixed, this can go into 4.10.0.

[danielplohmann]

Changelog, pin comment and the separate 0.6.0 entry all read right now. Approving for 4.10.0.