Release pipeline for v0.1.0

[Frikallo]

Release pipeline for v0.1.0

Closes #21 (the prebuilt-binaries + Xcode/CLT-doc parts).

What this adds

• scripts/package_release.sh — local-first packaging script. Configures with AXIOM_DIST_BUILD=ON (which statically links libomp/libgomp/OpenBLAS into libaxiom), builds, stages a bin/ + lib/ + include/ tree, fixes rpaths (install_name_tool on macOS, patchelf on Linux), and refuses to ship if otool -L/ldd shows a host-path leak. Designed to be runnable on a developer machine so packaging changes can be validated before pushing tags.
• .github/workflows/release.yml — matrix workflow (macos-14 / macos-13 / ubuntu-22.04). Each job calls the script. A final job creates the GitHub release and attaches all three tarballs. Triggers on v* tag push, plus workflow_dispatch for manual test runs (no release created on dispatch).
• README — new "Install" section pointing at releases with the xattr -dr com.apple.quarantine recipe; "Build from source" gains the one-line note that macOS source builds need full Xcode (not just CLT) because of xcrun metal / xcrun metallib. That covers ask (3) from the issue.
• .gitignore — adds build-release/ and dist/ (script outputs, never committed).

Tarball layout

parakeet-v0.1.0-<os>-<arch>/
├── bin/
│   ├── parakeet
│   └── example-server
├── lib/
│   ├── libaxiom.<ver>.{dylib,so}
├── include/parakeet/...
├── README.md
└── LICENSE

Runtime deps after dist build:

• macOS: Metal, Foundation, MPS, MPSGraph, libc++, libSystem (all system).
• Linux: libstdc++, libc, libpthread (all system).

The Metal .metallib is embedded into libaxiom.dylib via axiom's existing AXIOM_EMBED_METAL_LIBRARY=ON. Runtime uses newLibraryWithData:, which is part of the macOS Metal framework — no Xcode or CLT required on the user side.

What was validated locally (macos-arm64)

• Script ran clean.
• Tarball extracted to /tmp runs bin/parakeet --help (exit 0).
• otool -L shows no /opt/homebrew paths and no repo-absolute paths; @rpath/libaxiom.0.dylib resolves via the bundled lib.
• End-to-end: real model + real audio with --gpu produced the expected transcript ("Well, I don't wish to see it any more, observed Phoebe..."), encoder ran in 11ms — embedded .metallib worked through newLibraryWithData: from the relocated tarball.