Honor transform_source with the native parser

[Dextheking1]

Fixes #21222 (the transform_source half).

The issue reports two things broken under parallel checking. The shadow-file half was already handled — requires_read() checks the shadow map and works in threaded and --num-workers builds. The genuinely broken half is options.transform_source: the native parser path skipped it (it was only applied on the old-parser path), and the parallel build passed source=None, so the parser read the untransformed file from disk.

This applies the transform before dispatching to the native parser, and makes requires_read() force a source read when a transform is configured.

Tests: new TransformSourceSuite in mypy/test/testparse.py covering the native parser, the old parser, and a parallel build. The native-parser cases fail without the fix; all pass with it. Full testparse suite (249 passed, 75 skipped) and shadow-related testcheck (13 passed) green.

Disclosure: I used an LLM assistant while developing this fix.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[Dextheking1]

The only failing check is Test suite with py314t-ubuntu, mypyc-compiled, and the failure is in testConcurrentCircularNativeImports_multi (mypyc/test-data/run-multimodule.test), a free-threaded concurrency test: KeyError: 'other_b' while two threads import other_a/other_b concurrently under --nogil.

This is unrelated to the change here. The PR only touches mypy/build.py, mypy/parse.py, and mypy/test/testparse.py, none of which are involved in the mypyc native import path the test exercises (the test compiles its own modules with mypyc and imports them in a ThreadPoolExecutor). The failure is a race in the free-threaded import machinery, not in the transform code.

I checked: the py314t job is not in the required set for merging (the check's mergeStateStatus is UNSTABLE, not BLOCKED), and the change is mergeable otherwise. Happy to re-run if it flips back to a real failure.