Treat a missing parent package as "optional module unavailable"

[LeSingh1]

_optional_cuda_import() decides whether a ModuleNotFoundError means "the optional
module is not installed" (return None) or "the module is installed but one of its
dependencies is broken" (re-raise), by comparing err.name against the requested name:

    except ModuleNotFoundError as err:
        if err.name != fully_qualified_modname:
            raise
        return None

That comparison misses the most common case. ModuleNotFoundError.name is the
outermost missing name, so an absent parent package reports the parent:

>>> importlib.import_module("cuda.bindings.nvjitlink").__name__   # cuda-bindings not installed
ModuleNotFoundError: No module named 'cuda.bindings'      # err.name == "cuda.bindings"

Reproduced directly, with only cuda-pathfinder importable:

>>> from cuda.pathfinder._optional_cuda_import import _optional_cuda_import
>>> _optional_cuda_import("cuda.bindings.nvjitlink")
Traceback (most recent call last):
  ...
ModuleNotFoundError: No module named 'cuda.bindings'

The target module is exactly as unavailable as it would be if only the leaf were missing,
but the exception escapes instead of returning None.

Why it matters

Both callers today are in cuda.core, where cuda-bindings is an optional dependency
— cuda_core/pyproject.toml lists only cuda-pathfinder and numpy as required, with
cuda-bindings in the cu12/cu13 extras — and a None return selects a documented
fallback:

• cuda_core/cuda/core/_linker.pyx:687 → falls back to the driver cuLink* APIs
• cuda_core/cuda/core/_program.pyx:669 → falls back to NVRTC-only compilation

So the escaping exception defeats both fallbacks in precisely the configuration they exist
for: cuda-core installed without the bindings extra.

Fix

Accept a missing ancestor package as "unavailable" as well, via a small helper that
requires a real dotted-path boundary (startswith(missing + ".")), so a sibling that
merely shares a string prefix is still re-raised. The "don't mask a broken dependency"
half of the original intent is untouched.

Tests

Three added to cuda_pathfinder/tests/test_optional_cuda_import.py:

• ..._returns_none_when_parent_package_missing — stubs the exact ModuleNotFoundError
  CPython raises. Fails on main.
• ..._returns_none_for_a_really_uninstalled_package — no stub at all, goes through the
  real import machinery with a name that cannot exist. Fails on main.
• ..._reraises_for_a_string_prefix_that_is_not_an_ancestor — pins the dotted-boundary
  check (cuda.bindings is a string prefix of cuda.bindings_extra, not its parent).
  Passes on main too, on purpose.

The existing ..._reraises_nested_module_not_found test does not cover this case: it uses
an unrelated name (not_a_real_dependency), never an ancestor.

Verified: 2 of the 3 fail against upstream/main, all pass with the change, and the rest
of cuda_pathfinder/tests has the same pass/fail set as main. ruff check,
ruff format --check, and mypy (the pre-commit mypy-pathfinder invocation) are clean.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.