#1482 [Feature]: publish a ready-made Docker image, so users do not compile the engine themselves

Beschreibung

**What I want:** to `docker pull` a Strata image and start it, instead of building it myself. Today the Docker path in docs/INSTALL.md means every user compiles the engine on their own machine before the container ever starts. A GitHub Actions workflow that builds the image and pushes it to GHCR on a `v*` tag would fix that; the repo has no workflows today.

**The question first.** The Dockerfile says to build on the PC it runs on, and that's right for a local build. A published image can't follow that rule: it's built on a CI machine, and CI CPUs vary from run to run. So it must be built for the AVX2 floor instead — the same choice the ready-made engine already makes (CMakeLists.txt:31-34), and the same illegal-instruction failure #411 #412 #419 fixed for the vision encoder. A published image is the convenience build, not the fastest one. If you only want native builds, this issue can close.

**What it costs.** I tried it on my fork: [one uncached build](https://github.com/nullata/Strata/actions/runs/37717607320) took **1h32m**, the image is **5.3 GB**. GHCR needs no secrets and no setup beyond `packages: write`. The image is based on NVIDIA's CUDA devel image, so publishing it is a licensing call for you to make.

**Scope:** CUDA only, 64-bit x86 only, no test jobs (CI has no GPUs). Needs driver >=580; older-GPU users stay on setup.py's CUDA 12 path either way.

The code is on [nullata/Strata `docker-image-ci`](https://github.com/Niko1221/Strata/compare/main...nullata:Strata:docker-image-ci): one new workflow file, plus a `PORTABLE` build arg in the Dockerfile (default off, so local builds are unchanged). The test image is built and pushed, but I have not run it on a GPU yet. Happy to open a PR.