[alpae]

Adds Windows wheels (x86_64 and arm64) and Windows to the CI test matrix.

Fix

aligned_malloc in DynProgr_sse_double.c aligned its buffers by casting the pointer to long. On 64-bit Windows long is only 32 bits, so the pointer was truncated. That is a likely reason why the tests failed when Windows wheels were tried in 2022. It now uses uintptr_t.

Build

• wheels for windows-latest (x86_64) and windows-11-arm (arm64), built with MSVC and /openmp. delvewheel bundles the MSVC OpenMP runtime (vcomp140.dll)
• no 32-bit (win32) wheels
• no arm64 wheel for Python 3.10, as numpy has no Windows arm64 wheels for 3.10 (they start with numpy 2.3, which needs Python ≥ 3.11)
• ci: windows-latest in the test matrix; run steps use bash on all platforms, e.g. for the dist/*.tar.gz glob

Warnings

Cleaned up the MSVC warnings in our code:

• explicit casts where gap costs (double) are passed as char/short and lengths (Py_ssize_t) as int
• pointers in debug prints use %p instead of a cast to unsigned long
• c_align_strings no longer assigns the double tot in the loop initialiser of i

With this, the x86_64 build has no MSVC warnings left. On arm64, the only remaining warning comes from simde itself (simde/x86/sse.h, a uint64_t to int conversion), not from pyopa.

Results

All checks of this PR pass:

• wheels: every wheel passes the test suite (22 passed, 1 skipped), i.e. Windows x86_64 for Python 3.10–3.15, Windows arm64 for 3.11–3.15, and the existing Linux and macOS wheels
• ci: all 18 test jobs, including windows-latest with Python 3.10–3.15
• full-tests: the comparison with all 35,156 darwin reference results

The changelog has entries for the Windows wheels and the fixes.

[alpae]

should also fix #1