Move model download to runtime; keep image tiny for free tier

Switch to free cpu-basic: ubuntu base, drop CUDA/-ngl

Deploy llama.cpp web UI with Qwythos-9B MTP model