[CAOShurong]

Follow-up to #299 and ecf9675.

The AREF XY record stores three absolute points. The rectangular-array check
currently compares two cross coordinates with zero, so classification changes
when the same array is translated. For example, translating the file attached
to #299 makes its four instances overlap: offsets (0, 50, 100, 150) become
(0, 0, 0, 0).

Compare those coordinates with the corresponding origin coordinates instead.
This preserves the existing vector calculation and rotation/reflection handling;
raw integer equality also avoids subtraction overflow.

The synthetic regression covers rectangular, swapped, skewed, reversed and
zero vectors, translated origins, transforms, unit conversion and large legal
coordinates. It does not include the reporter's binary file.

Validation on Windows 11, CPython 3.13.1, MSVC:

• Independently built and non-editably installed unmodified main and the patch.
  The new regression has 14 failures / 52 passes on main and passes all 66
  cases repeatedly with the patch. All 91 original outcomes are unchanged;
  the full suite passes 157 tests.
• Ran python -m pytest <checkout>/tests/library_test.py -k aref_vectors -q
  and python -m pytest <checkout>/tests -q outside the source package, with
  fresh explicit --basetemp directories.
• Checked the original attachment and its translated variant through the
  installed API, plus GDS/OAS/streaming/flattened round-trips. KLayout 0.30.12
  independently compares polygon geometry for 68 inputs across four routes.
• Exercised the native C++ reader/exporters and missing-input handling; built
  the wheel/sdist and checked source/package/installed bytes.
• Changed-line C++ formatting and the new Python test's formatting pass.
  Whole-file Black retains one pre-existing formatting difference; compiler
  warnings match unmodified main.

Linux/macOS, other Python versions and production EDA signoff were not run
locally. No writer, tolerance or public API changes are included.

[heitzmann]

@CAOShurong Could you please take a look at the test errors? They seem to be just a matter of fixing the tolerance.

[CAOShurong]

Updated both vector checks to use the same coordinate-derived absolute tolerance as the offsets (rtol=0). The 157-test suite passes locally, and the 14 baseline regression failures are unchanged.