Build manylinux/musllinux aarch64 wheels

[colinbjohnson]

Replaces the duplicated manylinux_x86_64/musllinux_i686 matrix entries with aarch64 builds. Uses GitHub's native ubuntu-24.04-arm runner rather than QEMU emulation, so no cross-arch build step is needed.

Also bumps actions/checkout, cibuildwheel, and upload/download-artifact past the retired v3 actions — GitHub now hard-fails any job using actions/upload-artifact@v3, which was breaking every job in this workflow (not just aarch64) regardless of branch. upload-artifact@v4 requires a unique name per upload, so each matrix leg and the sdist job now upload under distinct names, merged back together in upload_pypi via download-artifact's pattern/merge-multiple.

Closes #80. Related to #76 (same duplicate-entry fix, predates native arm64 runner availability) and #77/#78 (same action-version bumps, needed for CI to run at all).

[colinbjohnson]

Closing for now while we finish validating this locally (no floating ubuntu-latest-arm runner alias exists yet — pinned to ubuntu-24.04-arm: https://github.blog/changelog/2024-06-24-github-actions-ubuntu-24-04-image-now-available-for-arm64-runners/). Will reopen once CI is green.

[colinbjohnson]

Reopening — Linux aarch64 (manylinux + musllinux) and macOS arm64 now build successfully on this branch.

macosx_x86_64 still fails independently of this change: macos-latest now points to Apple Silicon hardware, and build_parasail() in setup.py doesn't pass --host/ARCHFLAGS through to the bundled C library's configure, so it cross-builds the wrong architecture when CIBW_ARCHS: x86_64 is requested on an arm64 runner (delocate-wheel: Failed to find any binary with the required architecture: 'x86_64'). This would fail identically on unmodified master today — it's a separate, pre-existing issue unrelated to this PR's goal, leaving out of scope for now.