Ollama 0.30.0: llama.cpp integration, flash-attention default for Qwen/Gemma, broader model support. Firsthand upgrade notes, known issues to watch.