refactor(pypi): don't use Python to patch wheels

[aignas]

1. Refactored patch_whl.bzl — Extracted fix_record_content() as a public
   pure function (no rctx dependency) so it can be unit tested directly.

2. Unit tests (patch_whl_tests.bzl) — 5 new rules_testing tests for fix_record_content:

   • Adds missing files to RECORD
   • Returns None when nothing is missing
   • Preserves quoting style of existing entries
   • Skips excluded files (INSTALLER, RECORD*, REQUESTED)
   • Skips .whl files

3. Integration test (patch_whl_patch_test.py) — End-to-end test that:

• Creates a wheel from testdata/pkg/ via whl_from_dir_repo
• Applies a patch via whl_library(whl_patches=...)
• Verifies pkg.PATCHED is True (was False) and pkg.DATA is unchanged

The integration test follows the established pattern: test data in
tests/pypi/patch_whl/testdata/, registered in internal_dev_deps.bzl,
wired into MODULE.bazel, and built as a py_test. All 205 pypi tests
pass.

With this we can potentially more easily separate whl and sdist path
ways.

[gemini-code-assist]

Code Review

This pull request replaces the Python-dependent wheel patching mechanism with a native Starlark implementation that extracts, patches, fixes the RECORD file natively, and repacks the wheel using system utilities. Feedback on the changes highlights several issues: the CSV parsing logic for RECORD files does not handle filenames with commas or quotes correctly; both the Unix and Windows repacking commands will recursively include the output .whl file itself unless explicitly excluded; and the directory traversal loop uses an excessively large range of 10 million, which should be reduced to prevent potential memory and performance issues in Starlark.