llama.cpp and ggml hit v1.0.0 with semantic versioning, removed APIs, FP8 KV caches, a MoE cache and server hardening. What breaks and how to upgrade.