Build NIF 2.18 precompiled artifacts on OTP 29, release v0.8.22

[nyo16]

Problem

Installing llama_cpp_ex on OTP 29 (NIF 2.18) fails with a 404 when fetching the precompiled binary, even though mix.exs advertises NIF 2.18
support via make_precompiler_nif_versions.

Root cause

The precompile matrix built each target on OTP 27 and OTP 28, intending the OTP 28 jobs to produce the 2.18 artifacts — but OTP 28 still
reports NIF 2.17 (2.18 only arrived with OTP 29). Both jobs therefore produced identically named nif-2.17 tarballs, the second upload clobbered
the first, and no 2.18 artifact was ever published. The checksum job's --ignore-unavailable flag silently skipped the missing files, so every
release shipped this way unnoticed.

Retroactively uploading 2.18 assets to v0.8.21 wouldn't help: the published Hex package ships a checksum.exs containing only the 2.17 entries,
so a new release is required.

Changes

• precompile.yml — second matrix entry per target now builds on OTP 29 / Elixir 1.20 (verified available on builds.hex.pm for ubuntu-22.04 and
  erlef/otp_builds for macOS), producing real nif-2.18 artifacts alongside the OTP 27 nif-2.17 ones
• mix.exs — version bump to 0.8.22
• CHANGELOG.md — v0.8.22 entry (no llama.cpp submodule change)
• docs/release-guide.md — corrected the matrix description

Verification

• mix compile --warnings-as-errors and mix format --check-formatted pass
• Full test suite on OTP 29 / NIF 2.18 / Elixir 1.20 with local GGUF models: 147 passed, 1 skipped, 7 excluded