`PyImport_CreateModuleFromInitfunc()`: wrong `__name__` for submodules, non-ASCII names rejected, inittab name clashes

[itamaro]

Bug report

Bug description:

Since PyImport_CreateModuleFromInitfunc() was added (in 3.15, gh-116146), I ran into some issues from real world use internally at Meta (with 3.15+, as well as our 3.14 fork, which has this backported). Collected here as a single issue for convenience - let me know if splitting is preferred.

The API is built on the internal create_builtin() helper, and inherits a few inittab behaviors that don't make sense for an explicitly-passed init function.

1. Single-phase submodules get the short name. The package context is never set while initfunc runs, so a single-phase init that creates its module as sub (what pybind11's PYBIND11_MODULE(sub, m) does) ends up with __name__ == 'sub' while the spec says pkg.sub. The original impl handled that, but it was lost somewhere on the way.
2. Non-ASCII spec names raise UnicodeEncodeError, even for multi-phase modules, which the dynamic loader accepts.
3. Names registered in PyImport_Inittab are handled silently. Depending on the state of the builtin, initfunc is ignored, or the module created here replaces the builtin for later imports, or (multi-phase builtin) initfunc's module overwrites the real one in sys.modules. sys and builtins fall in the first bucket.
4. Undocumented semantics. Single-phase modules are added to sys.modules by the call itself, multi-phase ones aren't. The spec name is the module's identity, so a second call with the same name and a different initfunc returns the cached module and never calls the new function.

The existing test only covers one top-level single-phase and one top-level multi-phase module, so none of this was caught.

AI disclosure: I ran into the first issue in production, used Claude Opus 5 and Fable 5.1 to root cause, which flagged the other issues while investigating. Fable 5.1 prepared the fixes.

Demo output before any fix:

pkg.sub: repr=<module 'sub'> __name__=sub
non-ASCII name: UnicodeEncodeError: 'ascii' codec can't encode character '\xf6'
same name, initfunc B after A: which=A  calls_a=1 calls_b=0
'_random' via custom initfunc, then `import _random`: <module 'same'>  hasattr(Random)=False

Tests demonstrating all of the above: https://github.com/itamaro/cpython/tree/gh-116146-initfunc-tests-only

Plan

One PR each:

• Set the package context while calling initfunc, like the dynamic loader does (1)
• Accept non-ASCII names for multi-phase init; single-phase gets the same SystemError as dynamic loading (2)
• Raise ImportError for names registered in PyImport_Inittab (3)
• Document the sys.modules and name-identity behavior (4)

CPython versions tested on:

CPython main branch, 3.15

Operating systems tested on:

Linux, macOS

Linked PRs

• gh-157384: Set package context in PyImport_CreateModuleFromInitfunc() #157387
• gh-157384: Refuse inittab names in PyImport_CreateModuleFromInitfunc() #157388
• gh-157384: Document and test PyImport_CreateModuleFromInitfunc() semantics #157389
• gh-157384: Accept non-ASCII names in PyImport_CreateModuleFromInitfunc() #157390

[itamaro]

cc @encukou

All should be backported to 3.15, but not necessarily during the RC, 3.15.1 should be fine.