fix: upgrade to Q6_K (600MB) better quality

feat: switch to MiniCPM-V-4.6-Thinking GGUF

fix: use hf_hub_download at runtime like OpenWolf-Agent

fix: pin llama-cpp-python==0.3.23, disable mmap

fix: use create_completion instead of create_chat_completion, reduce ctx

fix: switch to MiniCPM3-4B (official openbmb GGUF)

fix: switch to Qwen2.5-1.5B for better llama.cpp compatibility

fix: use pre-compiled llama-cpp-python wheel + model in image

fix: bake GGUF model into Docker image during build

fix: use pre-compiled llama-server binary (zero compilation)

fix: use ninja-build + CMAKE_ARGS for llama-cpp-python build

init: MiniCPM-2B GGUF text Space