Compressed VLM inference from a single Containerfile — turboquant-vllm v1.1

Build a container with turboquant-vllm baked in, serve a vision-language model with 3.76x KV cache compression, and verify it works — one Containerfile, one flag.