fix: add cibuildwheel arch overrides to avoid host-native AVX-512 compilation

[XBeg9]

Summary

• Add cibuildwheel overrides to force specific architecture targets for manylinux builds
• x86_64: Enable Haswell target (-DENABLE_HASWELL=ON)
• aarch64: Enable ARMv8A target (-DENABLE_ARMV8A=ON)

This prevents the build from using host-native AVX-512 instructions that may not be available on all target machines.

Test plan

• Verify CI builds pass for manylinux x86_64
• Verify CI builds pass for manylinux aarch64

[CLAassistant]

All committers have signed the CLA.

[XBeg9]

Related to #101 but takes a different approach:

PR #101 adds a new X86_ARCH_NATIVE CMake flag that enables -march=native builds. This is useful for local development where you want to leverage all CPU features available on your machine. However, it also sets X86_ARCH_NATIVE="ON" as a default in pyproject.toml, which means wheel builds would use the CI machine's native instruction set—potentially including AVX-512 or other instructions not available on end-user machines.

This PR (#102) specifically targets cibuildwheel/CI builds by adding overrides that force safe, widely compatible architecture targets:

• x86_64: Haswell (AVX2, no AVX-512)
• aarch64: ARMv8A (baseline ARM64)

This ensures distributed wheels work on a broad range of hardware without illegal instruction errors (ref #92).

TL;DR: #101 is for local development flexibility; this PR is for ensuring distributed wheels have safe architecture targets in CI.