Fix setup.py cross-compilation flags for cibuildwheel

- macOS: detect x86_64 cross-compile via CIBW_ARCHS_MACOS, use
  -msse4.2 -mpopcnt instead of -mcpu=apple-m1 for x86_64 targets
- Linux x86_64: use -msse4.2 -mpopcnt instead of -march=native
  for manylinux portability
- Linux aarch64: use -march=armv8-a+simd

Previously, machine() returned 'arm64' on M1 runners even when
building x86_64 wheels, causing -mcpu=apple-m1 to be passed
incorrectly. And -march=native in manylinux containers didn't
include popcnt, breaking _mm_popcnt_u64 compilation.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: mrecachinas

setup.py

@@ -10,15 +10,22 @@
 from os import environ
 
 extra_compile_args = []
-if system().lower() == "darwin" and (
-    machine().lower() == "arm64" or environ.get("CIBW_ARCHS_MACOS", "") == "arm64"
-):
-    extra_compile_args.append("-mcpu=apple-m1")
+cibw_arch = environ.get("CIBW_ARCHS_MACOS", "")
+cibw_linux_arch = environ.get("CIBW_ARCHS_LINUX", "")
+host_machine = machine().lower()
+
+if system().lower() == "darwin":
+    if cibw_arch == "x86_64" or (cibw_arch == "" and host_machine != "arm64"):
+        extra_compile_args.extend(["-msse4.2", "-mpopcnt"])
+    else:
+        extra_compile_args.append("-mcpu=apple-m1")
 elif uname().system == "Windows":
     extra_compile_args.append("-O2")
     extra_compile_args.append("/d2FH4-")
+elif cibw_linux_arch == "aarch64" or host_machine == "aarch64":
+    extra_compile_args.append("-march=armv8-a+simd")
 else:
-    extra_compile_args.append("-march=native")
+    extra_compile_args.extend(["-msse4.2", "-mpopcnt"])
 
 setup(
     ext_modules=[