ci: consolidate llama-cpp-darwin into the matrix-driven Darwin flow

[mudler]

Summary

Move the bespoke llama-cpp-darwin + llama-cpp-darwin-publish top-level jobs in backend.yml into the matrix-driven Darwin flow that already handles all 34 other Darwin backends. The bespoke jobs were the only Darwin code path skipping the path filter — they rebuilt on every push/cron whether backend/cpp/llama-cpp/ was touched or not.

What changes

• .github/backend-matrix.yml — new entry under includeDarwin: { backend: llama-cpp, tag-suffix: -metal-darwin-arm64-llama-cpp, lang: go } (no build-type, the bespoke build script doesn't read one).
• .github/workflows/backend_build_darwin.yml — adds an if: inputs.backend == 'llama-cpp' branch in the build step that drives make backends/llama-cpp-darwin (the bespoke scripts/build/llama-cpp-darwin.sh compiles three CMake variants and bundles dylibs via otool — doesn't fit the generic build-darwin-go-backend mold). The llama-cpp-aware ccache setup blocks already in this file (lines 127, 134, 144) were the half-finished consolidation that this PR completes.
• scripts/changed-backends.js — inferBackendPathDarwin gains a special case so llama-cpp on Darwin maps to backend/cpp/llama-cpp/ (where the C++ sources live), not the non-existent backend/go/llama-cpp/ that the generic rule would produce.
• .github/workflows/backend.yml — deletes ~80 lines: the bespoke llama-cpp-darwin and llama-cpp-darwin-publish jobs are gone.
• Go version bump 1.24.x → 1.25.x in backend.yml and backend_pr.yml for the Darwin matrix. The bespoke job ran on 1.25.x, so this preserves llama-cpp's prior toolchain. The other 34 Darwin matrix backends pick this up too — none have a known reason to pin 1.24.

Behavior change

• Before: llama-cpp-darwin rebuilt on every backend.yml trigger (push, cron, tag).
• After: same path-filter as every other Darwin backend — only rebuilds when backend/cpp/llama-cpp/ changes (or via the weekly Sunday cron / a tag push, which both force the full matrix).
• All 34 existing Darwin backends now run on Go 1.25.x. Worth flagging in review.

Test plan

• PR-side backend_pr.yml runs generate-matrix cleanly with the new entry — already validated locally; smoke test confirms llama-cpp is in matrix-darwin output and the per-backend boolean llama-cpp=true lands when forced.
• After merge, a touch commit to backend/cpp/llama-cpp/ schedules backend-jobs-darwin (llama-cpp, ...) only — not the unconditional bespoke job (which is gone).
• The matrix entry produces the same final published tag (localai/localai-backends:master-metal-darwin-arm64-llama-cpp and the quay equivalent) — verified by reading docker/metadata-action config (unchanged).
• Go 1.25.x compiles all 34 other Darwin backends without regression.

Revert plan

If Go 1.25.x breaks something on Darwin, revert this commit; the bespoke jobs come back. If the matrix flow's make backends/llama-cpp-darwin invocation has a subtle env mismatch with the bespoke job's, the special-case build step is the focused unit of revert.

Assisted-by: Claude:claude-opus-4-7