fix(neon): under-aligned store in generic swizzle kernel

[vsaraikin]

Fixes #1260.

The generic neon swizzle kernel stores the batch into a local std::array<T, size>, which is only aligned to alignof(T) — 1 byte for the 8-bit batches that take this path (armv7 neon has no vqtbl1q; 8/16-bit swizzles fall through to the generic kernel). store_aligned requires the full batch alignment, so whenever the compiler places the array at a non-16-byte-aligned stack offset the alignment assertion fires: reduce_max → kernel::swizzle(neon) → store_aligned abort — exactly the Debian armhf backtrace in the issue.

This is not armhf-specific: the same abort reproduces on arm64 macOS with batch<unsigned char, xsimd::neon> compiled at -O2 with assertions enabled:

using batch_type = xsimd::batch<unsigned char, xsimd::neon>;
alignas(16) unsigned char raw[16] = { 0, 1, 2, 3, /* … */ };
auto b = batch_type::load_aligned(raw);
xsimd::reduce_max(b); // Assertion failed: store location is not properly aligned

The fix declares the buffer alignas(A::alignment()), matching the existing idiom in xsimd_neon64.hpp (alignas(A::alignment()) double data[] = …).

Adds a NEON-guarded regression test exercising the generic swizzle kernel and reduce_max for all four 8/16-bit batch types on the neon arch. One honest caveat: whether the assertion fires on the old code depends on the frame layout the compiler happens to pick (it fires in the issue's Debian build and in the -O2 repro above, but not in a local -O0 doctest build), so the test documents and exercises the path rather than deterministically aborting on regression.

Full test suite passes on arm64 macOS (391 test cases / 5928 assertions).

[DiamonDinoia]

no need to run this only in the neon case you can run this as a unit test on all architectures.

[vsaraikin]

Done — dropped the NEON guard; the test now iterates xsimd::supported_architectures (gated on Arch::available()), so the narrow swizzle kernels and reduce_max run for every enabled arch on each CI target.

[serge-sans-paille]

LGTM once @DiamonDinoia comment has been addressed.

[serge-sans-paille]

Thanks 🙇