#96 Add docker build

Description

The engine is compiled during docker build, so the first start only downloads the model.

docker build has no GPU, so the CUDA architectures are fixed by an ARG (80;86;89;120) instead of read from nvidia-smi; the engine is a fat binary with a cubin per arch. BUILD.json is written with source=local and the source hash, so setup reuses it instead of recompiling.

The entry point keeps the install config on the volume and selects it by family and model, so switching between models already set up needs no setup pass. REINSTALL stays for settings recorded inside a config.