fix(python): resolve annotations through TYPE_CHECKING imports (#1277)

[DeusData]

What

The import that an annotation's type comes from usually sits under a module-level if TYPE_CHECKING:, to
avoid an import cycle:

from typing import TYPE_CHECKING
if TYPE_CHECKING:
    from homeassistant.core import HomeAssistant

async def async_setup(hass: HomeAssistant) -> bool:
    hass.async_create_task(...)        # needs HomeAssistant to resolve

parse_python_imports walked only the module's direct statements, so that import was never recorded. Every
call through such an annotation fell to a name guess or got no edge at all.

• Extraction records the imports of a module-level if TYPE_CHECKING: / if <module>.TYPE_CHECKING:
  consequence block, flagged CBMImport.type_checking_only.
• The elif/else branches are runtime alternatives and stay unrecorded, as every other compound statement's
  imports do. A name such a branch binds again is bound at runtime, so its TYPE_CHECKING import is not an
  annotation-only import, and it stays unrecorded too. This is the compat shape, e.g.
  else: try: from airflow.sdk.log import mask_secret. A star import there covers every name of the
  block. Those names keep exactly the behaviour they have today.
• One rule for every CALLS import-map builder: cbm_import_binds_runtime_name(), which is false for a
  flagged import. So no CALLS import map binds an annotation-only name, and a runtime call through one gets
  no edge.
• The annotation consumers use the flagged imports: the cross-file base / field-type resolution and the
  Python LSP. The LSP binds them in an annotation-only frame, consulted by annotation resolution after the
  ordinary scope chain.

No index-format change.

Evidence

Base main 1346ac4 vs this change, 26 Python repos (manifest v1), production builds, macOS arm64,
--repeat 2, every repeat byte-identical.

type	base	this PR	+added	-removed
CALLS	1,355,020	1,371,605	16,611	26
TESTS	511,519	521,807	10,289	1
IMPORTS	477,476	486,103	8,627	0
CALL_REFERENCE	26,918	26,941	23	0
USAGE	1,672,278	1,672,256	0	22
INHERITS	78,598	78,597	0	1

• New CALLS (16,611): almost all LSP-resolved. The strategies are lsp_method 16,301,
  lsp_method_union 118, lsp_operator_dunder 111 and lsp_callable_alias 12, plus a few name
  strategies (86). The biggest share is home-assistant (14,430), whose integrations import HomeAssistant,
  ConfigEntry and friends under TYPE_CHECKING. Then scrapy 634, airflow 379, sqlalchemy 219, poetry 181,
  sentry 154 and mypy 136. TESTS follow from the new CALLS.
• New IMPORTS (8,627): the TYPE_CHECKING imports themselves, from the file to the imported module or
  symbol, as for runtime imports.
• Removed CALLS (26): 13 are retargeted on the same line to an LSP-typed target. The other 13 were wrong
  name guesses that are now gone: event.get_event_type() and dag_maker.create_dagrun() had been bound
  to same-named module functions, and an openlineage Dataset(...) to an unrelated task-sdk Dataset.
• USAGE -22: 16 of them come back as CALL_REFERENCE with the same source and target. The other 6 were
  cross-component name guesses.
• INHERITS -1: sentry's SentryPrinter(honcho.printer.Printer) has an external base, and the in-repo
  Printer it was bound to was wrong.

Judged precision of the additions

Seeded samples (seed 20261017), each judged against its call line and target definition: 50/50 correct.

• home-assistant CALLS: 20/20. 96% of its 14,430 new edges are hass.config_entries.<method>(...) reaching
  ConfigEntries.<method>, because HomeAssistant.config_entries is annotated with a type that core.py
  imports under TYPE_CHECKING. The rest go through hass.auth (AuthManager), hass.http
  (HomeAssistantHTTP), the recorder (Recorder) and annotated entity parameters. A spot check of that
  remainder (12 edges) is also correct.
• CALLS in the other 25 repos: 20/20. Annotated receivers reach their class's own or inherited method:
  ansible self._loader reaches DataLoader, scrapy self.stats reaches StatsCollector, mypy builder
  reaches IRBuilder / LowLevelIRBuilder, airflow dag_maker reaches DagMaker, sympy Expr reaches
  Basic methods, and so on.
• TESTS: 10/10. Each links a test function to the ConfigEntries method it calls.

Tests

• python_type_checking_imports_flagged_issue1277 (extraction): TYPE_CHECKING and typing.TYPE_CHECKING
  imports are recorded and flagged. A runtime import is not flagged, and an else branch and an ordinary
  if stay unrecorded.
• python_type_checking_rebound_in_else_unrecorded_issue1277 (extraction): a name the else branch binds
  again under try/except is not recorded. A star import in an elif covers every name, and a pure
  annotation import stays flagged.
• python_type_checking_typed_field_issue1277{,_parallel} (pipeline): the reporter's typed-field chain
  with both imports under TYPE_CHECKING resolves through the LSP. Controls: there is no import_map and no
  LSP CALLS edge through the annotation-only name.
• All RED without the fix.
• Affected suites (complexity, extraction, extraction_imports, grammar_imports, registry, pipeline,
  cross_repo, py_lsp, py_lsp_stress, py_lsp_scale, lsp_resolution_probe, edge_imports, edge_types_probe,
  convergence_probe, matrix_known_classes, parallel, incremental, integration) pass.

Part of #1277 (the TYPE_CHECKING half).

[DeusData]

Lint fix (3ead775): the memory-core linter failed because this change added a raw free in two
import-map builders (pass_calls.c, pass_lsp_cross.c). build_import_map now checks the predicate on
the name's span in the edge properties before copying it (cbm_imports_local_span_binds_runtime_name,
which cbm_imports_local_binds_runtime_name now wraps). The cross-LSP builder sends an annotation-only
name down its existing no-value path. Same behaviour, no new raw allocator site.
make -f Makefile.cbm lint-memory-core passes on the branch merged with main.

The graph cannot change from this commit, so the corpus measurement in the description still holds.