Add NVIDIA GPU image for script agent

[synesthesiam]

Summary

• add a CUDA-based Dockerfile.gpu for NVIDIA GPUs
• expose llama.cpp GPU layer offloading through --n-gpu-layers
• map agent CLI options to namespaced SCRIPT_AGENT_* environment variables through a container entrypoint
• retain standard Hugging Face and NVIDIA environment variables
• document Docker and Docker Compose usage with a LAN-reachable Home Assistant URL

Validation

• built the GPU image locally
• confirmed CUDA detected an NVIDIA GPU and llama.cpp reported GPU offload support
• verified all agent-owned container settings use the SCRIPT_AGENT_* prefix
• ran all 96 script-agent tests