Add an opt-in ABI3T build for aiscat

[honglei]

This is the implementation draft discussed in #620. It is intentionally
opt-in and additive: the version-specific wheels remain the preferred path.

The conversion loop does not fork in this version. convert_object,
convert_value, and wrap_annotated remain the single semantic
implementation. The two performance-sensitive ABI differences are isolated in
small helpers:

• version-specific builds use _PyDict_NewPresized; ABI3T uses PyDict_New;
• version-specific builds use PyList_SET_ITEM; ABI3T uses the checked
  PyList_SetItem call.

The larger conditional section is outside the conversion loop: ABI3T needs a
negative-basicsize heap type, PyObject_GetTypeData, and the
PyModExport__core module export hook because PyObject_HEAD is opaque.

For maintenance, I would treat the shared conversion semantics as the sole
reference implementation. The version-specific build is the performance
baseline, with ABI-specific mechanics kept at the narrow boundaries above.
Both variants run the same decode tests, including annotated output and all
supported input object types.

The build uses scikit-build-core 1.0.3's ABI3T signal as the single source for
the source ABI, Stable-ABI linkage, suffix, and combined wheel tag. A
free-threaded Python 3.15 build produces
cp315-abi3.abi3t / _core.abi3t.so; an inconsistent classic-ABI request is
rejected rather than producing a mislabeled wheel. The ABI3T command requests
CMake 3.26+ explicitly, while ordinary version-specific builds retain their
existing CMake 3.15 floor.

Current validation

• ABI3T wheel passed the project tests on Python 3.15.0rc1 free-threaded with
  the GIL still disabled, and on a GIL-enabled Python 3.15 build.
• The cp315t and cp313 version-specific wheels passed the same tests.
• 100,000 Decoder create/decode/destroy cycles passed on both 3.15 runtimes.
• The Linux extension exports PyModExport__core, not PyInit__core; all 45
  imported CPython symbols are in the CPython 3.15.0rc1 Stable ABI manifest.
• On 5,000,000 real AIS input lines, with file I/O and grouping excluded, the
  ABI3T feed() median was 6.58% below cp315t at one thread and 5.04% below at
  four threads. This supports retaining cp315t as the preferred artifact.

This is not merge-ready release plumbing. An actual Windows wheel build/import
(including DLL dependency inspection), ABI3T-aware auditing once available,
and a separately controlled CI/release job remain open gates.