perf: reuse parsers across parallel graph builds

[Lum1104]

Linked issue

Related to #189 and the parser-load hardening merged in #459.

What & why

Parallel full and incremental builds called CodeParser(...) once per file. That discarded the parser instance cache on every task and repeatedly paid language-pack/parser initialization inside each worker. On a mixed-language checkout of this repository, the default 8-process raw build took 36.23 seconds even though serial parsing took 7.71 seconds.

This change keeps one parser per worker and repository in threading.local() storage:

• process workers reuse parser state across their assigned files;
• thread workers retain isolated mutable parser state;
• changing repository roots invalidates the worker-local parser;
• executor selection, worker-count policy, error handling, and serial SQLite writes remain unchanged.

The worker-local design also applies to incremental builds because both build paths share _parse_single_file.

Impact and benchmark

Same 215-file mixed-language repository, fresh project-local data directory, --skip-postprocess, Python 3.14.2 on a 10-core Apple Silicon host:

Configuration	Before	After
8 process workers (default)	36.23 s	7.11 s
4 process workers	5.59 s	6.32 s

The default path improves by about 5.1x. The two post-change worker counts produced identical graphs: 4,420 nodes and 34,162 edges. I intentionally left worker-count heuristics unchanged; the measured regression is parser lifecycle, while a global worker cap needs broader repository-size and platform data.

How it was tested

uv run pytest tests/test_incremental.py::TestParseExecutorSelection tests/test_incremental.py::TestParallelParsing tests/test_mcp_stdio_shutdown.py -q
# 10 passed

CRG_PARSE_EXECUTOR=thread uv run pytest tests/test_incremental.py::TestParallelParsing::test_parallel_build_produces_same_results -q
CRG_PARSE_EXECUTOR=process uv run pytest tests/test_incremental.py::TestParallelParsing::test_parallel_build_produces_same_results -q
# 1 passed for each executor

uv tool run ruff check code_review_graph/incremental.py tests/test_incremental.py
# All checks passed

uv tool run mypy code_review_graph/incremental.py --ignore-missing-imports --no-strict-optional
# Success: no issues found

uv run pytest -q
# 1963 passed, 5 skipped, 2 xpassed, 1 unrelated failure

The full-suite failure is pre-existing on Python 3.14/macOS: test_windows_server_still_prewarms_before_mcp_run monkeypatches asyncio.WindowsSelectorEventLoopPolicy after setting sys.platform to win32, which raises inside Python 3.14 asyncio before this code path is reached.

Checklist

• Tests added for new functionality
• All tests pass: one unrelated Python 3.14 platform-emulation failure documented above
• Linting passes
• Type checking passes for the changed production module
• Lines are at most 100 characters
• Docstring and implementation comments updated for the worker lifecycle behavior