Native ARM64/NEON support (via SIMDe)

[BenjaminDEMAILLE]

Native ARM64/NEON support (via SIMDe)

This makes hifiasm build and run natively on aarch64/arm64 (Linux ARM64 and
Apple Silicon) in addition to x86, with the x86 build byte-for-byte unchanged.

Fixes #288. Supersedes #641: that PR used sse2neon and built fine, but this one
uses SIMDe instead, which also covers
AVX (not just SSE), is header-complete for the intrinsics hifiasm uses, and comes
with a cross-arch CI matrix that proves the x86 and ARM64 assemblies are identical.

Why the current build fails on ARM

The Makefile passes -msse4.2 -mpopcnt (the latter is a hard error on aarch64),
and Levenshtein_distance.h / ksw2_extz2_sse.c include x86 intrinsic headers
(<emmintrin.h>, <smmintrin.h>, <nmmintrin.h>, <immintrin.h>) that do not
exist on ARM.

What changed

All source changes are gated behind defined(__aarch64__) || defined(__arm__),
so on x86 the preprocessor takes the original path and codegen is unchanged.

• Makefile: detect the target with uname -m. On aarch64/arm64, drop
  -msse4.2 -mpopcnt, use -march=armv8-a+simd, and add -Ithird_party/simde.
  ARCH and ARCH_FLAGS remain overridable, e.g.
  make ARCH_FLAGS="-mcpu=neoverse-v2" or make ARCH=aarch64. The x86 branch is
  identical to before (-msse4.2 -mpopcnt, empty INCLUDES).
• Levenshtein_distance.h: on ARM, include the SIMDe x86/sse4.2.h and
  x86/avx2.h headers instead of the four x86 headers. One subtlety: GCC's x86
  _mm_srli_epi32 accepts a runtime shift count, but the NEON translation
  requires a compile-time constant, so the two variable-count uses go through a
  tiny HA_SRLI_EPI32_VAR macro that expands to the identical _mm_srli_epi32
  on x86 and to the variable-count shift on ARM.
• ksw2_extz2_sse.c: provide the SSE2/SSE4.1 kernel via SIMDe and define
  __SSE2__/__SSE4_1__ so the body compiles on ARM. Note this file is not in
  the Makefile OBJS and its only caller is commented out, so it stays unlinked;
  it is ported (and verified to compile standalone on ARM) so it builds cleanly
  if it is ever re-enabled, with no change to either binary.
• third_party/simde/: vendored SIMDe, pinned to v0.8.2
  (commit 71fd833d9666141edcd1d3c109a80e228303d8d7, MIT). Only the simde/
  header tree and COPYING are included (no tests/docs), so a plain
  git clone && make works with no submodule step.
• .github/workflows/ci.yaml + test/gen_hifi_reads.py: CI now builds a
  matrix of Linux x86_64, Linux aarch64, macOS x86_64, and macOS arm64, and a
  smoke job assembles a tiny deterministic synthetic HiFi set on both Linux arches
  and asserts the x86_64 and aarch64 graphs are identical.

Notes on scope

No popcnt intrinsic (_mm_popcnt* / __builtin_popcount*) is used anywhere, so
dropping -mpopcnt on ARM needs no code change. And although <immintrin.h> was
included, no _mm256_*/AVX intrinsic is actually called; SIMDe's avx2.h is
still included on ARM to preserve the original include surface.

x86 is unchanged

• Every source edit is inside an ARM #if; the #else branch is the original code.
• The Makefile x86 branch reproduces the original flags and empty INCLUDES.
• ksw2_extz2_sse.c stays out of OBJS, so the link line is unchanged on all arches.

Testing

make succeeds and produces a working binary on all four targets:

Target	Result
Linux x86_64	CI: build + smoke assembly pass
Linux aarch64	CI: build + smoke assembly pass
macOS arm64	CI + local (Apple Silicon): build + smoke pass
macOS x86_64	local cross-compile passes (identical -msse4.2 -mpopcnt path, no SIMDe)

Determinism (the assembled graph is identical across architectures) is checked
two independent ways, both producing a byte-identical 67,868 B primary-contig GFA:

• CI, Linux, GCC: x86_64 vs aarch64 (see the cross-arch graph determinism job).
• Local, macOS, clang: x86_64 (under Rosetta) vs native arm64.

Locally on Apple Silicon, hifiasm -o test -t2 reads.fq.gz on a small synthetic
HiFi set assembles a single 59,609 bp contig, and repeated runs are byte-identical.

CI run (on my fork): https://github.com/BenjaminDEMAILLE/hifiasm/actions/runs/29027889900
The three Linux/macOS-arm64 build legs, both smoke legs, and the determinism job
are green; the macOS x86_64 leg was still waiting on a (scarce) Intel runner at
the time of writing and is additionally covered by the local cross-compile above.