iOS: build the _multiprocessing extension (importable, not spawnable)

[ndonkoHenri]

Build the _multiprocessing C-extension for iOS. CPython marks it n/a on iOS (alongside _posixsubprocess and _posixshmem) because process spawning can't work in the app sandbox. But the extension itself builds fine on Darwin — macOS ships it — and import multiprocessing[.connection/.synchronize] is perfectly usable. Today, any package that imports multiprocessing at module load crashes on iOS with ModuleNotFoundError: No module named '_multiprocessing' — e.g. scikit-learn's sklearn.callback._transport, which does an unconditional from multiprocessing.connection import ....

This makes multiprocessing importable, not runnable. Process().start(), Pool, etc. still fail — that's the inherent iOS sandbox limit, and _posixsubprocess/_posixshmem stay disabled. The goal is only to stop the import-time crash, which is what actually bites in practice.

How

Two edits in darwin/build_ios.py, both after the patch and before Apple build:

1. Flip the module on in the generated configure: py_cv_module__multiprocessing=n/a → =yes. Version-agnostic — it hits both the vendored ios_patches/<3.14 configure and upstream 3.14+'s PY_STDLIB_MOD_SET_NA; a no-op if the string isn't present.
2. Add ac_cv_func_sem_timedwait=no and ac_cv_func_sem_clockwait=no to the existing ios-config.site (next to the pipe2/dup3 overrides). Darwin has neither; the SDK may still declare them, so this forces _multiprocessing/semaphore.c onto its no-timeout fallback instead of referencing symbols that won't link.

Validation (3.12.13)

• Build: produces _multiprocessing.xcframework with both ios-arm64 and ios-arm64_x86_64-simulator slices (a proper _multiprocessing.framework each) — compiles cleanly through Darwin's missing sem_timedwait.
• Runtime (iOS simulator): injected the built framework into a Flet app and ran a probe.
  • Before: _multiprocessing MISSING, 1/5 — only the lazy import multiprocessing survives; .connection, .synchronize, SemLock all fail.
  • After: _multiprocessing PRESENT, 5/5 — including a real SemLock acquire / release / trywait round-trip (so sem_open works and the no-sem_timedwait fallback is fine), plus the .connection / .synchronize imports packages actually need.

Useful for scikit-learn on iOS (flet-dev/mobile-forge#92) without a per-recipe workaround, and fixes the same import crash for every other package that merely imports multiprocessing.