docker: add xz support and validate backend extraction

[fl0rianr]

Summary

Adds xz-utils to the container runtime image so Lemonade can extract .tar.xz backend archives.

Also extends the container smoke test with a real .tar.xz extraction check and temporarily uploads the ARM64 image artifact on pull request runs for renewed hardware validation.

Why

The first real ARM64 container test successfully verified:

• linux/arm64 image architecture,
• CLI startup,
• server startup,
• /live health check.

Installing the llamacpp:cuda backend then failed with:

tar (child): xz: Cannot exec: No such file or directory

The downloaded backend archive was valid, but the runtime image contained tar without the external xz decompressor required for .tar.xz files.

Changes

• Add xz-utils to the Docker runtime dependencies.
• Add a .tar.xz create-and-extract smoke test for AMD64 and ARM64 images.
• Temporarily upload the ARM64 image artifact on pull request runs so the backend installation can be retested on real ARM64 hardware.
• Keep the existing manual artifact upload option unchanged.

Testing

• AMD64 container build passes
• ARM64 container build passes
• CLI smoke tests pass
• .tar.xz extraction passes on AMD64
• .tar.xz extraction passes on ARM64
• Server /live health checks pass
• Updated ARM64 artifact installs llamacpp:cuda successfully on real hardware

[fl0rianr]

@betonme Thanks for the proper run test you did afterwards. Would you be able to test this artifact again? The updated ARM64 image with xz-utils is ready here:

https://github.com/lemonade-sdk/lemonade/actions/runs/29166996672

Under Artifacts, download:

lemonade-server-arm64-c0e8a5fedbd15ca03ca390c70cbcbc860d85a6c9

This build also passed the new .tar.xz extraction smoke test. Please retry installing llamacpp:cuda with this image, and also a small interference test would be nice.

For your info: Currently Lemonade uses CUDA 12.8 but we are planning to update towards CUDA 12.9 which brings better support for sm121, but this testing should still work with the current version.

[betonme]

Thanks for preparing, I'll check it.

[betonme]

@fl0rianr
It works fine.
I downloaded and installed llama:cuda and stablediffusion:cuda
A quick test with Llama3.2 and SD-Turbo shows both runs on the GPU.
Thanks