v1.1.0: CPU-safe GPU detection, real GPU offload, hardened fallible calls, doctor command

fix(space): install libgomp1 in runtime; llama.cpp links OpenMP

fix(space): bookworm base (gcc 12) and coherent ggml flags

fix(space): compile llama-cpp-python in a builder stage (no manylinux wheel exists)

THOX mesh node: OpenAI endpoint + mesh self-registration