fix(python): walk a class nested in a class under its outer class

[DeusData]

What

The Python LSP walked a class nested in a class body (plain or decorated) as if it were declared at module
level. Its methods were walked under proj.m.Inner while extraction defines them under
proj.m.Outer.Inner, so:

• calls inside those methods never joined their extracted callers, and
• self.method(), super().method() and builtin calls inside a nested class fell back to name guesses
  (suffix_match) that pick an unrelated same-named method elsewhere in the project.

py_process_class_in (py_lsp.c) now walks a nested class, and a decorated nested class, under the outer
class's QN, which is the QN extraction gives it. Classes nested in functions (#1548) and top-level classes
are unchanged.

Stacked on #2690 (decorated nested classes are extracted). This PR's base branch is #2690's branch,
fix/py-decorated-nested-class; once #2690 merges, this PR's diff against main is this change alone. No
index-format change.

Evidence

Base #2690 @4d71327c vs this change @ace384c9, 26 Python repos (manifest v1), production builds, macOS
arm64, --repeat 2, every repeat byte-identical. The corpus delta was measured against #2690 at 4d71327,
before #2679 (dotted base QNs, a different py_lsp.c function) reached main. The branch has since merged
main, and PR CI tests the merged tree. Nodes are unchanged, and so is every edge type except:

type	base	this PR	+added	-removed
CALLS	1,355,032	1,355,095	173	110
TESTS	511,529	511,531	5	3

Another 250 CALLS keep their target and move from a name guess to an LSP strategy.

• Every one of the 110 removed CALLS is a suffix_match guess. Each one was replaced, on the same call
  line, by an LSP-resolved edge to a different target. The 173 added CALLS are all LSP-resolved
  (lsp_method 112, lsp_builtin_method 28, lsp_builtin_constructor 15, lsp_super 8, others 10).
• Read per repo, every replacement swaps a wrong guess for the right binding:
  • mitmproxy (64): the kaitai-generated nested classes (DtlsClientHello.CipherSuites, .Cookie, ...)
    call their own _read, not Alpn._read.
  • sqlalchemy (17): HSTORE.Comparator.contains / has_all / ... reach the inherited
    TypeEngine.Comparator.operate, not associationproxy's operate.
  • wagtail (9): FieldPanel.BoundPanel / InlinePanel.BoundPanel self.is_shown / id_for_label /
    js_opts reach their own methods.
  • airflow (6): ReferenceModels.BaseDeadlineReference.evaluate_with's self._evaluate_with and
    SerializedBaseDeadlineReference.__eq__ / __hash__'s self.serialize_reference reach the class's
    own methods, not a subclass override.
  • home-assistant (5): Config._ConfigStore.async_save's super() call reaches
    helpers.storage.Store.async_save. A test-local TestFlow.SubentryFlowHandler calls its own
    async_step_*, not broadlink's BroadlinkFlowHandler.async_step_finish.
  • Smaller cases (sentry 4, matplotlib 3, django 1, sympy 1): super() reaches the real base (django
    UUIDField.get_db_prep_value), list() / dict() / str() reach the builtins, and
    ArrowStyle._Base.__call__'s self.transmute reaches _Base.transmute.
• TESTS: the 3 removed edges came from the wrong home-assistant guess above. The 5 added edges are
  celery smoke tests in nested test classes reaching canvas.group.apply_async and Signature.delay.

Tests

• pipeline_python_nested_class_methods_resolve{,_parallel}: calls inside the methods of a class nested
  in a class resolve through the LSP to the right targets, on the sequential and the parallel pipeline.
  RED without the fix (ASSERT(typed)).
• Affected suites (complexity, extraction, registry, pipeline, cross_repo, py_lsp, py_lsp_stress,
  py_lsp_scale, lsp_resolution_probe, incremental) pass on the merged tree.