Rebase to Alpine 3.24, use distro onnxruntime for the discovery embedding runtime

[IrosTheBeggar]

---

• I have read the contributing guideline and understand that I have made the correct modifications

---

Description:

Closes #35.

• Rebase to Alpine 3.24 (both Dockerfiles).
• Add gcompat and onnxruntime to the runtime packages.
• After npm install, replace the bundled glibc libonnxruntime.so.1 inside onnxruntime-node with a symlink to the distro's musl build, and remove the bundled CUDA/TensorRT provider libraries (~316 MB) that CPU inference never loads.
• Changelog entry in readme-vars.yml.

Benefits of this PR and context:

mStream ≥6.16 ships a discovery/recommendation engine whose embedding model runs on onnxruntime-node. That npm package bundles glibc-only binaries, so on this image the worker dies at import (Error loading shared library ld-linux-x86-64.so.2) and the feature can never build its data (#35).

gcompat alone is not enough — the bundled runtime references fortified glibc symbols (__vsnprintf_chk, __memcpy_chk, …) that the shim doesn't implement. But Alpine 3.24's community repo now packages a native musl onnxruntime 1.24.4, the same 1.24 line the npm binding targets. With the distro runtime symlinked in, the only glibc code left is the npm package's thin 384 KB N-API binding, which gcompat handles fine.

Net effect: the recommendation features work, and the image shrinks by ~350 MB (the deleted CUDA/TensorRT libs outweigh the added packages).

The if [ -d … ] guard keeps the build working even if a future mStream release drops the optional onnxruntime-node dependency. If mStream ever moves to a newer onnxruntime line before Alpine does, mStream detects the failed load (as of IrosTheBeggar/mStream#718, in the next release), logs one clear error, and disables the embedding pass — the image itself is never broken by drift.

How Has This Been Tested?

• Built this Dockerfile (x86_64) locally and ran it with collectDiscoveryData enabled against a small music library: the embedding pass completed real inferences — Discovery-embedding pass complete: 6 embedded, 0 error(s) — and the discovery DB contains 1280-d embeddings + genre predictions.
• The recipe was first validated clean-room in a plain alpine:3.24 container with the real model (18 MB EffNet ONNX): full inference passes with the system runtime + gcompat; fails without.
• Confirmed the onnxruntime apk exists for aarch64 at 3.24 as well (onnxruntime-1.24.4-r1 via qemu); the arm64 image itself was not run here and deserves a smoke in your CI.

Source / References:

• Issue: [FEAT] Migrating to a distro with glibc support #35
• Alpine package: https://pkgs.alpinelinux.org/package/v3.24/community/x86_64/onnxruntime
• mStream-side graceful degradation + docs: fix(discovery): musl/Alpine onnxruntime — degrade once loudly, glibc container path IrosTheBeggar/mStream#718