Automatic multi-arch Docker images in CI

[jaywonchung]

NVIDIA's Grace CPU is ARM, which means eventually we can expect people from benefiting from a native ARM Docker image. Hence multi-platform images is the way forward, but when I tested it last time, building was painfully slow (probably due to QEMU and cross-compilation) and it exceeded the storage capacity of the GitHub Actions runner. But now that we significantly reduced the size of the image by switching to NVIDIA/CUDA base images, I think we finally have good chances of getting it working.