fix: expose() skipped providers inherited from a parent Group

[lesnik512]

Why

_collect_fixtures discovered providers by walking vars(group) — that is
cls.__dict__, the class's own body only. A provider inherited from a parent
Group was silently skipped, so the fixture set disagreed with the container:

modern-di sees:               ['base_repo', 'child_repo']   # Child.get_named_providers()
expose installed (before):    ['child_repo']
container resolves base_repo: Repo(label='real')

The user gets base_repo registered and resolvable from the container, but no
base_repo fixture, with no error to say so.

This was drift, not a decision. The vars() walk dates to the initial commit
(2026-05-24); modern-di added the public Group.get_named_providers() a month
later (modern-python/modern-di#242, released in 2.20.0), which returns exactly
the name -> provider mapping this function built by hand, walking the full
MRO. The later seam-adoption pass (589674d) reviewed resolve_dependency and
add_providers and adopted the first; it never considered the group walk,
though the method already existed. Nothing in planning/decisions/ or
deferred.md records the own-body-only behaviour as intended.

Design

_collect_fixtures now asks each group for its named providers instead of
walking attributes itself. The cross-group collision check and the
raise-before-returning contract are unchanged. The local skip-non-Provider
branch is gone — get_named_providers() already filters to AbstractProvider,
and name shadowing now follows MRO order, so a child redefining a parent's
provider name yields the child's.

No dependency bump: the floor is already modern-di>=3,<4, well above the
2.20.0 that first shipped the method.

The README's expose() paragraph — "one pytest fixture per Provider class
attribute" — becomes true as written, so it needs no edit.

Compatibility

This can only add fixtures. The one visible edge: a newly-surfaced inherited
name can now collide across groups and raise the existing ValueError, where
before it was invisible.

Non-goals

• The inject/register wording in the README and in factory.py belongs to
  the in-flight vocabulary migration on docs/migrate-off-planning; the error
  string is left byte-identical here so the two do not conflict.
• The README does not mention that expose() with no groups raises
  TypeError. Real but separate; not fixed here.
• No planning/changes/ file. By the convention live on main this is the
  Lightweight lane and would want one, but docs/migrate-off-planning deletes
  that directory and is rebased and ready, so adding a file there buys a
  guaranteed conflict. The spec is this PR body instead, matching where the repo
  is going. Say the word and I'll add the change file.

Verification

just lint-ci clean. just test-ci and just test-branch pass at 100%
line and branch coverage, 17 tests. The new invariant test fails on the old
vars() walk with exactly the three missing inherited names.